"""Client for interacting with the LangSmith API.
Use the client to customize API keys / workspace ocnnections, SSl certs,
etc. for tracing.
Also used to create, read, update, and delete LangSmith resources
such as runs (~trace spans), datasets, examples (~records),
feedback (~metrics), projects (tracer sessions/groups), etc.
For detailed API documentation, visit: https://docs.smith.langchain.com/.
"""
from __future__ import annotations
import atexit
import collections
import concurrent.futures as cf
import contextlib
import datetime
import functools
import importlib
import importlib.metadata
import io
import itertools
import json
import logging
import os
import random
import threading
import time
import traceback
import typing
import uuid
import warnings
import weakref
from inspect import signature
from pathlib import Path
from queue import PriorityQueue
from typing import (
TYPE_CHECKING,
Any,
AsyncIterable,
Callable,
DefaultDict,
Dict,
Iterable,
Iterator,
List,
Literal,
Mapping,
Optional,
Sequence,
Tuple,
Type,
Union,
cast,
)
from urllib import parse as urllib_parse
import requests
from requests import adapters as requests_adapters
from requests_toolbelt import ( # type: ignore[import-untyped]
multipart as rqtb_multipart,
)
from typing_extensions import TypeGuard, overload
from urllib3.poolmanager import PoolKey # type: ignore[attr-defined, import-untyped]
from urllib3.util import Retry # type: ignore[import-untyped]
import langsmith
from langsmith import env as ls_env
from langsmith import schemas as ls_schemas
from langsmith import utils as ls_utils
from langsmith._internal import _orjson
from langsmith._internal._background_thread import (
TracingQueueItem,
)
from langsmith._internal._background_thread import (
tracing_control_thread_func as _tracing_control_thread_func,
)
from langsmith._internal._background_thread import (
tracing_control_thread_func_compress_parallel as _tracing_control_thread_func_compress_parallel,
)
from langsmith._internal._beta_decorator import warn_beta
from langsmith._internal._compressed_runs import CompressedRuns
from langsmith._internal._constants import (
_AUTO_SCALE_UP_NTHREADS_LIMIT,
_BLOCKSIZE_BYTES,
_BOUNDARY,
_SIZE_LIMIT_BYTES,
)
from langsmith._internal._multipart import (
MultipartPart,
MultipartPartsAndContext,
join_multipart_parts_and_context,
)
from langsmith._internal._operations import (
SerializedFeedbackOperation,
SerializedRunOperation,
combine_serialized_queue_operations,
compress_multipart_parts_and_context,
serialize_feedback_dict,
serialize_run_dict,
serialized_feedback_operation_to_multipart_parts_and_context,
serialized_run_operation_to_multipart_parts_and_context,
)
from langsmith._internal._serde import dumps_json as _dumps_json
try:
from zoneinfo import ZoneInfo # type: ignore[import-not-found]
except ImportError:
class ZoneInfo: # type: ignore[no-redef]
"""Introduced in python 3.9."""
if TYPE_CHECKING:
import pandas as pd # type: ignore
from langchain_core.runnables import Runnable
from langsmith import schemas
from langsmith.evaluation import evaluator as ls_evaluator
from langsmith.evaluation._arunner import (
AEVALUATOR_T,
ATARGET_T,
AsyncExperimentResults,
)
from langsmith.evaluation._runner import (
COMPARATIVE_EVALUATOR_T,
DATA_T,
EVALUATOR_T,
EXPERIMENT_T,
SUMMARY_EVALUATOR_T,
TARGET_T,
ComparativeExperimentResults,
ExperimentResults,
)
logger = logging.getLogger(__name__)
_urllib3_logger = logging.getLogger("urllib3.connectionpool")
X_API_KEY = "x-api-key"
WARNED_ATTACHMENTS = False
EMPTY_SEQ: tuple[Dict, ...] = ()
URLLIB3_SUPPORTS_BLOCKSIZE = "key_blocksize" in signature(PoolKey).parameters
def _parse_token_or_url(
url_or_token: Union[str, uuid.UUID],
api_url: str,
num_parts: int = 2,
kind: str = "dataset",
) -> Tuple[str, str]:
"""Parse a public dataset URL or share token."""
try:
if isinstance(url_or_token, uuid.UUID) or uuid.UUID(url_or_token):
return api_url, str(url_or_token)
except ValueError:
pass
# Then it's a URL
parsed_url = urllib_parse.urlparse(str(url_or_token))
# Extract the UUID from the path
path_parts = parsed_url.path.split("/")
if len(path_parts) >= num_parts:
token_uuid = path_parts[-num_parts]
_as_uuid(token_uuid, var="token parts")
else:
raise ls_utils.LangSmithUserError(f"Invalid public {kind} URL: {url_or_token}")
if parsed_url.netloc == "smith.langchain.com":
api_url = "https://api.smith.langchain.com"
elif parsed_url.netloc == "beta.smith.langchain.com":
api_url = "https://beta.api.smith.langchain.com"
return api_url, token_uuid
def _is_langchain_hosted(url: str) -> bool:
"""Check if the URL is langchain hosted.
Args:
url (str): The URL to check.
Returns:
bool: True if the URL is langchain hosted, False otherwise.
"""
try:
netloc = urllib_parse.urlsplit(url).netloc.split(":")[0]
return netloc.endswith("langchain.com")
except Exception:
return False
ID_TYPE = Union[uuid.UUID, str]
RUN_TYPE_T = Literal[
"tool", "chain", "llm", "retriever", "embedding", "prompt", "parser"
]
def _default_retry_config() -> Retry:
"""Get the default retry configuration.
If urllib3 version is 1.26 or greater, retry on all methods.
Returns:
Retry: The default retry configuration.
"""
retry_params = dict(
total=3,
status_forcelist=[502, 503, 504, 408, 425],
backoff_factor=0.5,
# Sadly urllib3 1.x doesn't support backoff_jitter
raise_on_redirect=False,
raise_on_status=False,
respect_retry_after_header=True,
)
# the `allowed_methods` keyword is not available in urllib3 < 1.26
# check to see if urllib3 version is 1.26 or greater
urllib3_version = importlib.metadata.version("urllib3")
use_allowed_methods = tuple(map(int, urllib3_version.split("."))) >= (1, 26)
if use_allowed_methods:
# Retry on all methods
retry_params["allowed_methods"] = None
return ls_utils.LangSmithRetry(**retry_params) # type: ignore
def close_session(session: requests.Session) -> None:
"""Close the session.
Args:
session (requests.Session): The session to close.
"""
logger.debug("Closing Client.session")
session.close()
def _validate_api_key_if_hosted(api_url: str, api_key: Optional[str]) -> None:
"""Verify API key is provided if url not localhost.
Args:
api_url (str): The API URL.
api_key (Optional[str]): The API key.
Returns:
None
Raises:
LangSmithUserError: If the API key is not provided when using the hosted service.
"""
# If the domain is langchain.com, raise error if no api_key
if not api_key:
if _is_langchain_hosted(api_url):
warnings.warn(
"API key must be provided when using hosted LangSmith API",
ls_utils.LangSmithMissingAPIKeyWarning,
)
def _get_tracing_sampling_rate() -> float | None:
"""Get the tracing sampling rate.
Returns:
Optional[float]: The tracing sampling rate.
"""
sampling_rate_str = ls_utils.get_env_var("TRACING_SAMPLING_RATE")
if sampling_rate_str is None:
return None
sampling_rate = float(sampling_rate_str)
if sampling_rate < 0 or sampling_rate > 1:
raise ls_utils.LangSmithUserError(
"LANGSMITH_TRACING_SAMPLING_RATE must be between 0 and 1 if set."
f" Got: {sampling_rate}"
)
return sampling_rate
def _get_write_api_urls(_write_api_urls: Optional[Dict[str, str]]) -> Dict[str, str]:
_write_api_urls = _write_api_urls or json.loads(
os.getenv("LANGSMITH_RUNS_ENDPOINTS", "{}")
)
processed_write_api_urls = {}
for url, api_key in _write_api_urls.items():
processed_url = url.strip()
if not processed_url:
raise ls_utils.LangSmithUserError(
"LangSmith runs API URL within LANGSMITH_RUNS_ENDPOINTS cannot be empty"
)
processed_url = processed_url.strip().strip('"').strip("'").rstrip("/")
processed_api_key = api_key.strip().strip('"').strip("'")
_validate_api_key_if_hosted(processed_url, processed_api_key)
processed_write_api_urls[processed_url] = processed_api_key
return processed_write_api_urls
def _as_uuid(value: ID_TYPE, var: Optional[str] = None) -> uuid.UUID:
try:
return uuid.UUID(value) if not isinstance(value, uuid.UUID) else value
except ValueError as e:
var = var or "value"
raise ls_utils.LangSmithUserError(
f"{var} must be a valid UUID or UUID string. Got {value}"
) from e
@typing.overload
def _ensure_uuid(value: Optional[Union[str, uuid.UUID]]) -> uuid.UUID: ...
@typing.overload
def _ensure_uuid(
value: Optional[Union[str, uuid.UUID]], *, accept_null: bool = True
) -> Optional[uuid.UUID]: ...
def _ensure_uuid(value: Optional[Union[str, uuid.UUID]], *, accept_null: bool = False):
if value is None:
if accept_null:
return None
return uuid.uuid4()
return _as_uuid(value)
@functools.lru_cache(maxsize=1)
def _parse_url(url):
parsed_url = urllib_parse.urlparse(url)
host = parsed_url.netloc.split(":")[0]
return host
class _LangSmithHttpAdapter(requests_adapters.HTTPAdapter):
__attrs__ = [
"max_retries",
"config",
"_pool_connections",
"_pool_maxsize",
"_pool_block",
"_blocksize",
]
def __init__(
self,
pool_connections: int = requests_adapters.DEFAULT_POOLSIZE,
pool_maxsize: int = requests_adapters.DEFAULT_POOLSIZE,
max_retries: Union[Retry, int, None] = requests_adapters.DEFAULT_RETRIES,
pool_block: bool = requests_adapters.DEFAULT_POOLBLOCK,
blocksize: int = 16384, # default from urllib3.BaseHTTPSConnection
) -> None:
self._blocksize = blocksize
super().__init__(pool_connections, pool_maxsize, max_retries, pool_block)
def init_poolmanager(self, connections, maxsize, block=False, **pool_kwargs):
if URLLIB3_SUPPORTS_BLOCKSIZE:
# urllib3 before 2.0 doesn't support blocksize
pool_kwargs["blocksize"] = self._blocksize
return super().init_poolmanager(connections, maxsize, block, **pool_kwargs)
class Client:
"""Client for interacting with the LangSmith API."""
__slots__ = [
"__weakref__",
"api_url",
"api_key",
"retry_config",
"timeout_ms",
"session",
"_get_data_type_cached",
"_web_url",
"_tenant_id",
"tracing_sample_rate",
"_filtered_post_uuids",
"tracing_queue",
"_anonymizer",
"_hide_inputs",
"_hide_outputs",
"_info",
"_write_api_urls",
"_settings",
"_manual_cleanup",
"_pyo3_client",
"compressed_runs",
"_data_available_event",
"_futures",
]
def __init__(
self,
api_url: Optional[str] = None,
*,
api_key: Optional[str] = None,
retry_config: Optional[Retry] = None,
timeout_ms: Optional[Union[int, Tuple[int, int]]] = None,
web_url: Optional[str] = None,
session: Optional[requests.Session] = None,
auto_batch_tracing: bool = True,
anonymizer: Optional[Callable[[dict], dict]] = None,
hide_inputs: Optional[Union[Callable[[dict], dict], bool]] = None,
hide_outputs: Optional[Union[Callable[[dict], dict], bool]] = None,
info: Optional[Union[dict, ls_schemas.LangSmithInfo]] = None,
api_urls: Optional[Dict[str, str]] = None,
) -> None:
"""Initialize a Client instance.
Args:
api_url (Optional[str]): URL for the LangSmith API. Defaults to the LANGCHAIN_ENDPOINT
environment variable or https://api.smith.langchain.com if not set.
api_key (Optional[str]): API key for the LangSmith API. Defaults to the LANGCHAIN_API_KEY
environment variable.
retry_config (Optional[Retry]): Retry configuration for the HTTPAdapter.
timeout_ms (Optional[Union[int, Tuple[int, int]]]): Timeout for the HTTPAdapter. Can also be a 2-tuple of
(connect timeout, read timeout) to set them separately.
web_url (Optional[str]): URL for the LangSmith web app. Default is auto-inferred from
the ENDPOINT.
session (Optional[requests.Session]): The session to use for requests. If None, a new session will be
created.
auto_batch_tracing (bool, default=True): Whether to automatically batch tracing.
anonymizer (Optional[Callable[[dict], dict]]): A function applied for masking serialized run inputs and outputs,
before sending to the API.
hide_inputs (Optional[Union[Callable[[dict], dict], bool]]): Whether to hide run inputs when tracing with this client.
If True, hides the entire inputs. If a function, applied to
all run inputs when creating runs.
hide_outputs (Optional[Union[Callable[[dict], dict], bool]]): Whether to hide run outputs when tracing with this client.
If True, hides the entire outputs. If a function, applied to
all run outputs when creating runs.
info (Optional[ls_schemas.LangSmithInfo]): The information about the LangSmith API.
If not provided, it will be fetched from the API.
api_urls (Optional[Dict[str, str]]): A dictionary of write API URLs and their corresponding API keys.
Useful for multi-tenant setups. Data is only read from the first
URL in the dictionary. However, ONLY Runs are written (POST and PATCH)
to all URLs in the dictionary. Feedback, sessions, datasets, examples,
annotation queues and evaluation results are only written to the first.
Raises:
LangSmithUserError: If the API key is not provided when using the hosted service.
If both api_url and api_urls are provided.
"""
if api_url and api_urls:
raise ls_utils.LangSmithUserError(
"You cannot provide both api_url and api_urls."
)
if (
os.getenv("LANGSMITH_ENDPOINT") or os.getenv("LANGCHAIN_ENDPOINT")
) and os.getenv("LANGSMITH_RUNS_ENDPOINTS"):
raise ls_utils.LangSmithUserError(
"You cannot provide both LANGSMITH_ENDPOINT / LANGCHAIN_ENDPOINT "
"and LANGSMITH_RUNS_ENDPOINTS."
)
self.tracing_sample_rate = _get_tracing_sampling_rate()
self._filtered_post_uuids: set[uuid.UUID] = set()
self._write_api_urls: Mapping[str, Optional[str]] = _get_write_api_urls(
api_urls
)
if self._write_api_urls:
self.api_url = next(iter(self._write_api_urls))
self.api_key: Optional[str] = self._write_api_urls[self.api_url]
else:
self.api_url = ls_utils.get_api_url(api_url)
self.api_key = ls_utils.get_api_key(api_key)
_validate_api_key_if_hosted(self.api_url, self.api_key)
self._write_api_urls = {self.api_url: self.api_key}
self.retry_config = retry_config or _default_retry_config()
self.timeout_ms = (
(timeout_ms, timeout_ms)
if isinstance(timeout_ms, int)
else (timeout_ms or (10_000, 90_001))
)
self._web_url = web_url
self._tenant_id: Optional[uuid.UUID] = None
# Create a session and register a finalizer to close it
session_ = session if session else requests.Session()
self.session = session_
if ls_utils.get_env_var("USE_RUN_COMPRESSION"):
self._futures: set[cf.Future] = set()
self.compressed_runs: Optional[CompressedRuns] = CompressedRuns()
self._data_available_event = threading.Event()
else:
self.compressed_runs = None
self._info = (
info
if info is None or isinstance(info, ls_schemas.LangSmithInfo)
else ls_schemas.LangSmithInfo(**info)
)
weakref.finalize(self, close_session, self.session)
atexit.register(close_session, session_)
# Initialize auto batching
if auto_batch_tracing and self.compressed_runs is not None:
self.tracing_queue: Optional[PriorityQueue] = None
threading.Thread(
target=_tracing_control_thread_func_compress_parallel,
# arg must be a weakref to self to avoid the Thread object
# preventing garbage collection of the Client object
args=(weakref.ref(self),),
).start()
elif auto_batch_tracing:
self.tracing_queue = PriorityQueue()
threading.Thread(
target=_tracing_control_thread_func,
# arg must be a weakref to self to avoid the Thread object
# preventing garbage collection of the Client object
args=(weakref.ref(self),),
).start()
else:
self.tracing_queue = None
# Mount the HTTPAdapter with the retry configuration.
adapter = _LangSmithHttpAdapter(
max_retries=self.retry_config,
blocksize=_BLOCKSIZE_BYTES,
# We need to set the pool_maxsize to a value greater than the
# number of threads used for batch tracing, plus 1 for other
# requests.
pool_maxsize=_AUTO_SCALE_UP_NTHREADS_LIMIT + 1,
)
self.session.mount("http://", adapter)
self.session.mount("https://", adapter)
self._get_data_type_cached = functools.lru_cache(maxsize=10)(
self._get_data_type
)
self._anonymizer = anonymizer
self._hide_inputs = (
hide_inputs
if hide_inputs is not None
else ls_utils.get_env_var("HIDE_INPUTS") == "true"
)
self._hide_outputs = (
hide_outputs
if hide_outputs is not None
else ls_utils.get_env_var("HIDE_OUTPUTS") == "true"
)
# To trigger this code, set the `LANGSMITH_USE_PYO3_CLIENT` env var to any value.
self._pyo3_client = None
if ls_utils.get_env_var("USE_PYO3_CLIENT") is not None:
langsmith_pyo3 = None
try:
import langsmith_pyo3 # type: ignore[import-not-found, no-redef]
except ImportError as e:
logger.warning(
"Failed to import `langsmith_pyo3` when PyO3 client was requested, "
"falling back to Python impl: %s",
repr(e),
)
if langsmith_pyo3:
# TODO: tweak these constants as needed
queue_capacity = 1_000_000
batch_size = 100
batch_timeout_millis = 1000
worker_threads = 1
try:
self._pyo3_client = langsmith_pyo3.BlockingTracingClient(
self.api_url,
self.api_key,
queue_capacity,
batch_size,
batch_timeout_millis,
worker_threads,
)
except Exception as e:
logger.warning(
"Failed to instantiate `langsmith_pyo3.BlockingTracingClient` "
"when PyO3 client was requested, falling back to Python impl: %s",
repr(e),
)
self._settings: Union[ls_schemas.LangSmithSettings, None] = None
self._manual_cleanup = False
def _repr_html_(self) -> str:
"""Return an HTML representation of the instance with a link to the URL.
Returns:
str: The HTML representation of the instance.
"""
link = self._host_url
return f'LangSmith Client'
def __repr__(self) -> str:
"""Return a string representation of the instance with a link to the URL.
Returns:
str: The string representation of the instance.
"""
return f"Client (API URL: {self.api_url})"
@property
def _host(self) -> str:
return _parse_url(self.api_url)
@property
def _host_url(self) -> str:
"""The web host url."""
return ls_utils.get_host_url(self._web_url, self.api_url)
@property
def _headers(self) -> Dict[str, str]:
"""Get the headers for the API request.
Returns:
Dict[str, str]: The headers for the API request.
"""
headers = {
"User-Agent": f"langsmith-py/{langsmith.__version__}",
"Accept": "application/json",
}
if self.api_key:
headers[X_API_KEY] = self.api_key
return headers
@property
def info(self) -> ls_schemas.LangSmithInfo:
"""Get the information about the LangSmith API.
Returns:
ls_schemas.LangSmithInfo: The information about the LangSmith API, or None if the API is
not available.
"""
if self._info is None:
try:
response = self.request_with_retries(
"GET",
"/info",
headers={"Accept": "application/json"},
timeout=(self.timeout_ms[0] / 1000, self.timeout_ms[1] / 1000),
)
ls_utils.raise_for_status_with_text(response)
self._info = ls_schemas.LangSmithInfo(**response.json())
except BaseException as e:
logger.warning(
f"Failed to get info from {self.api_url}: {repr(e)}",
)
self._info = ls_schemas.LangSmithInfo()
return self._info
def _get_settings(self) -> ls_schemas.LangSmithSettings:
"""Get the settings for the current tenant.
Returns:
dict: The settings for the current tenant.
"""
if self._settings is None:
response = self.request_with_retries("GET", "/settings")
ls_utils.raise_for_status_with_text(response)
self._settings = ls_schemas.LangSmithSettings(**response.json())
return self._settings
def _content_above_size(self, content_length: Optional[int]) -> Optional[str]:
if content_length is None or self._info is None:
return None
info = cast(ls_schemas.LangSmithInfo, self._info)
bic = info.batch_ingest_config
if not bic:
return None
size_limit = bic.get("size_limit_bytes")
if size_limit is None:
return None
if content_length > size_limit:
return (
f"The content length of {content_length} bytes exceeds the "
f"maximum size limit of {size_limit} bytes."
)
return None
def request_with_retries(
self,
/,
method: Literal["GET", "POST", "PUT", "PATCH", "DELETE"],
pathname: str,
*,
request_kwargs: Optional[Mapping] = None,
stop_after_attempt: int = 1,
retry_on: Optional[Sequence[Type[BaseException]]] = None,
to_ignore: Optional[Sequence[Type[BaseException]]] = None,
handle_response: Optional[Callable[[requests.Response, int], Any]] = None,
_context: str = "",
**kwargs: Any,
) -> requests.Response:
"""Send a request with retries.
Args:
method (str): The HTTP request method.
pathname (str): The pathname of the request URL. Will be appended to the API URL.
request_kwargs (Mapping): Additional request parameters.
stop_after_attempt (int, default=1): The number of attempts to make.
retry_on (Optional[Sequence[Type[BaseException]]]): The exceptions to retry on. In addition to:
[LangSmithConnectionError, LangSmithAPIError].
to_ignore (Optional[Sequence[Type[BaseException]]]): The exceptions to ignore / pass on.
handle_response (Optional[Callable[[requests.Response, int], Any]]): A function to handle the response and return whether to continue retrying.
_context (str, default=""): The context of the request.
**kwargs (Any): Additional keyword arguments to pass to the request.
Returns:
requests.Response: The response object.
Raises:
LangSmithAPIError: If a server error occurs.
LangSmithUserError: If the request fails.
LangSmithConnectionError: If a connection error occurs.
LangSmithError: If the request fails.
"""
request_kwargs = request_kwargs or {}
request_kwargs = {
"timeout": (self.timeout_ms[0] / 1000, self.timeout_ms[1] / 1000),
**request_kwargs,
**kwargs,
"headers": {
**self._headers,
**request_kwargs.get("headers", {}),
**kwargs.get("headers", {}),
},
}
if (
method != "GET"
and "data" in request_kwargs
and "files" not in request_kwargs
and not request_kwargs["headers"].get("Content-Type")
):
request_kwargs["headers"]["Content-Type"] = "application/json"
logging_filters = [
ls_utils.FilterLangSmithRetry(),
ls_utils.FilterPoolFullWarning(host=str(self._host)),
]
retry_on_: Tuple[Type[BaseException], ...] = (
*(retry_on or ()),
*(
ls_utils.LangSmithConnectionError,
ls_utils.LangSmithRequestTimeout, # 408
ls_utils.LangSmithAPIError, # 500
),
)
to_ignore_: Tuple[Type[BaseException], ...] = (*(to_ignore or ()),)
response = None
for idx in range(stop_after_attempt):
try:
try:
with ls_utils.filter_logs(_urllib3_logger, logging_filters):
response = self.session.request(
method,
(
self.api_url + pathname
if not pathname.startswith("http")
else pathname
),
stream=False,
**request_kwargs,
)
ls_utils.raise_for_status_with_text(response)
return response
except requests.exceptions.ReadTimeout as e:
logger.debug("Passing on exception %s", e)
if idx + 1 == stop_after_attempt:
raise
sleep_time = 2**idx + (random.random() * 0.5)
time.sleep(sleep_time)
continue
except requests.HTTPError as e:
if response is not None:
if handle_response is not None:
if idx + 1 < stop_after_attempt:
should_continue = handle_response(response, idx + 1)
if should_continue:
continue
if response.status_code == 500:
raise ls_utils.LangSmithAPIError(
f"Server error caused failure to {method}"
f" {pathname} in"
f" LangSmith API. {repr(e)}"
f"{_context}"
)
elif response.status_code == 408:
raise ls_utils.LangSmithRequestTimeout(
f"Client took too long to send request to {method}"
f"{pathname} {_context}"
)
elif response.status_code == 429:
raise ls_utils.LangSmithRateLimitError(
f"Rate limit exceeded for {pathname}. {repr(e)}"
f"{_context}"
)
elif response.status_code == 401:
raise ls_utils.LangSmithAuthError(
f"Authentication failed for {pathname}. {repr(e)}"
f"{_context}"
)
elif response.status_code == 404:
raise ls_utils.LangSmithNotFoundError(
f"Resource not found for {pathname}. {repr(e)}"
f"{_context}"
)
elif response.status_code == 409:
raise ls_utils.LangSmithConflictError(
f"Conflict for {pathname}. {repr(e)}" f"{_context}"
)
else:
raise ls_utils.LangSmithError(
f"Failed to {method} {pathname} in LangSmith"
f" API. {repr(e)}"
)
else:
raise ls_utils.LangSmithUserError(
f"Failed to {method} {pathname} in LangSmith API."
f" {repr(e)}"
)
except requests.ConnectionError as e:
recommendation = (
"Please confirm your LANGCHAIN_ENDPOINT."
if self.api_url != "https://api.smith.langchain.com"
else "Please confirm your internet connection."
)
try:
content_length = int(
str(e.request.headers.get("Content-Length"))
if e.request
else ""
)
size_rec = self._content_above_size(content_length)
if size_rec:
recommendation = size_rec
except ValueError:
content_length = None
api_key = (
e.request.headers.get("x-api-key") or "" if e.request else ""
)
prefix, suffix = api_key[:5], api_key[-2:]
filler = "*" * (max(0, len(api_key) - 7))
masked_api_key = f"{prefix}{filler}{suffix}"
raise ls_utils.LangSmithConnectionError(
f"Connection error caused failure to {method} {pathname}"
f" in LangSmith API. {recommendation}"
f" {repr(e)}"
f"\nContent-Length: {content_length}"
f"\nAPI Key: {masked_api_key}"
f"{_context}"
) from e
except Exception as e:
args = list(e.args)
msg = args[1] if len(args) > 1 else ""
msg = msg.replace("session", "session (project)")
if args:
emsg = "\n".join(
[str(args[0])]
+ [msg]
+ [str(arg) for arg in (args[2:] if len(args) > 2 else [])]
)
else:
emsg = msg
raise ls_utils.LangSmithError(
f"Failed to {method} {pathname} in LangSmith API. {emsg}"
f"{_context}"
) from e
except to_ignore_ as e:
if response is not None:
logger.debug("Passing on exception %s", e)
return response
except ls_utils.LangSmithRateLimitError:
if idx + 1 == stop_after_attempt:
raise
if response is not None:
try:
retry_after = float(response.headers.get("retry-after", "30"))
except Exception as e:
logger.warning(
"Invalid retry-after header: %s",
repr(e),
)
retry_after = 30
# Add exponential backoff
retry_after = retry_after * 2**idx + random.random()
time.sleep(retry_after)
except retry_on_:
# Handle other exceptions more immediately
if idx + 1 == stop_after_attempt:
raise
sleep_time = 2**idx + (random.random() * 0.5)
time.sleep(sleep_time)
continue
# Else we still raise an error
raise ls_utils.LangSmithError(
f"Failed to {method} {pathname} in LangSmith API."
)
def _get_paginated_list(
self, path: str, *, params: Optional[dict] = None
) -> Iterator[dict]:
"""Get a paginated list of items.
Args:
path (str): The path of the request URL.
params (Optional[dict]): The query parameters.
Yields:
The items in the paginated list.
"""
params_ = params.copy() if params else {}
offset = params_.get("offset", 0)
params_["limit"] = params_.get("limit", 100)
while True:
params_["offset"] = offset
response = self.request_with_retries(
"GET",
path,
params=params_,
)
items = response.json()
if not items:
break
yield from items
if len(items) < params_["limit"]:
# offset and limit isn't respected if we're
# querying for specific values
break
offset += len(items)
def _get_cursor_paginated_list(
self,
path: str,
*,
body: Optional[dict] = None,
request_method: Literal["GET", "POST"] = "POST",
data_key: str = "runs",
) -> Iterator[dict]:
"""Get a cursor paginated list of items.
Args:
path (str): The path of the request URL.
body (Optional[dict]): The query body.
request_method (Literal["GET", "POST"], default="POST"): The HTTP request method.
data_key (str, default="runs"): The key in the response body that contains the items.
Yields:
The items in the paginated list.
"""
params_ = body.copy() if body else {}
while True:
response = self.request_with_retries(
request_method,
path,
request_kwargs={
"data": _dumps_json(params_),
},
)
response_body = response.json()
if not response_body:
break
if not response_body.get(data_key):
break
yield from response_body[data_key]
cursors = response_body.get("cursors")
if not cursors:
break
if not cursors.get("next"):
break
params_["cursor"] = cursors["next"]
def upload_dataframe(
self,
df: pd.DataFrame,
name: str,
input_keys: Sequence[str],
output_keys: Sequence[str],
*,
description: Optional[str] = None,
data_type: Optional[ls_schemas.DataType] = ls_schemas.DataType.kv,
) -> ls_schemas.Dataset:
"""Upload a dataframe as individual examples to the LangSmith API.
Args:
df (pd.DataFrame): The dataframe to upload.
name (str): The name of the dataset.
input_keys (Sequence[str]): The input keys.
output_keys (Sequence[str]): The output keys.
description (Optional[str]): The description of the dataset.
data_type (Optional[DataType]): The data type of the dataset.
Returns:
Dataset: The uploaded dataset.
Raises:
ValueError: If the csv_file is not a string or tuple.
Examples:
.. code-block:: python
from langsmith import Client
import os
import pandas as pd
client = Client()
df = pd.read_parquet("path/to/your/myfile.parquet")
input_keys = ["column1", "column2"] # replace with your input column names
output_keys = ["output1", "output2"] # replace with your output column names
dataset = client.upload_dataframe(
df=df,
input_keys=input_keys,
output_keys=output_keys,
name="My Parquet Dataset",
description="Dataset created from a parquet file",
data_type="kv", # The default
)
"""
csv_file = io.BytesIO()
df.to_csv(csv_file, index=False)
csv_file.seek(0)
return self.upload_csv(
("data.csv", csv_file),
input_keys=input_keys,
output_keys=output_keys,
description=description,
name=name,
data_type=data_type,
)
def upload_csv(
self,
csv_file: Union[str, Tuple[str, io.BytesIO]],
input_keys: Sequence[str],
output_keys: Sequence[str],
*,
name: Optional[str] = None,
description: Optional[str] = None,
data_type: Optional[ls_schemas.DataType] = ls_schemas.DataType.kv,
) -> ls_schemas.Dataset:
"""Upload a CSV file to the LangSmith API.
Args:
csv_file (Union[str, Tuple[str, io.BytesIO]]): The CSV file to upload. If a string, it should be the path
If a tuple, it should be a tuple containing the filename
and a BytesIO object.
input_keys (Sequence[str]): The input keys.
output_keys (Sequence[str]): The output keys.
name (Optional[str]): The name of the dataset.
description (Optional[str]): The description of the dataset.
data_type (Optional[ls_schemas.DataType]): The data type of the dataset.
Returns:
Dataset: The uploaded dataset.
Raises:
ValueError: If the csv_file is not a string or tuple.
Examples:
.. code-block:: python
from langsmith import Client
import os
client = Client()
csv_file = "path/to/your/myfile.csv"
input_keys = ["column1", "column2"] # replace with your input column names
output_keys = ["output1", "output2"] # replace with your output column names
dataset = client.upload_csv(
csv_file=csv_file,
input_keys=input_keys,
output_keys=output_keys,
name="My CSV Dataset",
description="Dataset created from a CSV file",
data_type="kv", # The default
)
"""
data = {
"input_keys": input_keys,
"output_keys": output_keys,
}
if name:
data["name"] = name
if description:
data["description"] = description
if data_type:
data["data_type"] = ls_utils.get_enum_value(data_type)
data["id"] = str(uuid.uuid4())
if isinstance(csv_file, str):
with open(csv_file, "rb") as f:
file_ = {"file": f}
response = self.request_with_retries(
"POST",
"/datasets/upload",
data=data,
files=file_,
)
elif isinstance(csv_file, tuple):
response = self.request_with_retries(
"POST",
"/datasets/upload",
data=data,
files={"file": csv_file},
)
else:
raise ValueError("csv_file must be a string or tuple")
ls_utils.raise_for_status_with_text(response)
result = response.json()
# TODO: Make this more robust server-side
if "detail" in result and "already exists" in result["detail"]:
file_name = csv_file if isinstance(csv_file, str) else csv_file[0]
file_name = file_name.split("/")[-1]
raise ValueError(f"Dataset {file_name} already exists")
return ls_schemas.Dataset(
**result,
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
def _run_transform(
self,
run: Union[ls_schemas.Run, dict, ls_schemas.RunLikeDict],
update: bool = False,
copy: bool = False,
) -> dict:
"""Transform the given run object into a dictionary representation.
Args:
run (Union[ls_schemas.Run, dict]): The run object to transform.
update (Optional[bool]): Whether the payload is for an "update" event.
copy (Optional[bool]): Whether to deepcopy run inputs/outputs.
Returns:
dict: The transformed run object as a dictionary.
"""
global WARNED_ATTACHMENTS
if hasattr(run, "dict") and callable(getattr(run, "dict")):
run_create: dict = run.dict() # type: ignore
else:
run_create = cast(dict, run)
if "id" not in run_create:
run_create["id"] = uuid.uuid4()
elif isinstance(run_create["id"], str):
run_create["id"] = uuid.UUID(run_create["id"])
if "inputs" in run_create and run_create["inputs"] is not None:
if copy:
run_create["inputs"] = ls_utils.deepish_copy(run_create["inputs"])
run_create["inputs"] = self._hide_run_inputs(run_create["inputs"])
if "outputs" in run_create and run_create["outputs"] is not None:
if copy:
run_create["outputs"] = ls_utils.deepish_copy(run_create["outputs"])
run_create["outputs"] = self._hide_run_outputs(run_create["outputs"])
if not update and not run_create.get("start_time"):
run_create["start_time"] = datetime.datetime.now(datetime.timezone.utc)
# Only retain LLM & Prompt manifests
if "serialized" in run_create:
if run_create.get("run_type") not in (
"llm",
"prompt",
):
# Drop completely
run_create.pop("serialized", None)
elif run_create.get("serialized"):
# Drop graph
run_create["serialized"].pop("graph", None)
return run_create
@staticmethod
def _insert_runtime_env(runs: Sequence[dict]) -> None:
runtime_env = ls_env.get_runtime_environment()
for run_create in runs:
run_extra = cast(dict, run_create.setdefault("extra", {}))
# update runtime
runtime: dict = run_extra.setdefault("runtime", {})
run_extra["runtime"] = {**runtime_env, **runtime}
# update metadata
metadata: dict = run_extra.setdefault("metadata", {})
langchain_metadata = ls_env.get_langchain_env_var_metadata()
metadata.update(
{k: v for k, v in langchain_metadata.items() if k not in metadata}
)
def _filter_for_sampling(
self, runs: Iterable[dict], *, patch: bool = False
) -> list[dict]:
if self.tracing_sample_rate is None:
return list(runs)
if patch:
sampled = []
for run in runs:
run_id = _as_uuid(run["id"])
if run_id not in self._filtered_post_uuids:
sampled.append(run)
else:
self._filtered_post_uuids.remove(run_id)
return sampled
else:
sampled = []
for run in runs:
if (
# Child run
run["id"] != run.get("trace_id")
# Whose trace is included
and run.get("trace_id") not in self._filtered_post_uuids
# Or a root that's randomly sampled
) or random.random() < self.tracing_sample_rate:
sampled.append(run)
else:
self._filtered_post_uuids.add(_as_uuid(run["id"]))
return sampled
def create_run(
self,
name: str,
inputs: Dict[str, Any],
run_type: RUN_TYPE_T,
*,
project_name: Optional[str] = None,
revision_id: Optional[str] = None,
**kwargs: Any,
) -> None:
"""Persist a run to the LangSmith API.
Args:
name (str): The name of the run.
inputs (Dict[str, Any]): The input values for the run.
run_type (str): The type of the run, such as tool, chain, llm, retriever,
embedding, prompt, or parser.
project_name (Optional[str]): The project name of the run.
revision_id (Optional[Union[UUID, str]]): The revision ID of the run.
**kwargs (Any): Additional keyword arguments.
Returns:
None
Raises:
LangSmithUserError: If the API key is not provided when using the hosted service.
Examples:
.. code-block:: python
from langsmith import Client
import datetime
from uuid import uuid4
client = Client()
run_id = uuid4()
client.create_run(
id=run_id,
project_name=project_name,
name="test_run",
run_type="llm",
inputs={"prompt": "hello world"},
outputs={"generation": "hi there"},
start_time=datetime.datetime.now(datetime.timezone.utc),
end_time=datetime.datetime.now(datetime.timezone.utc),
hide_inputs=True,
hide_outputs=True,
)
"""
project_name = project_name or kwargs.pop(
"session_name",
# if the project is not provided, use the environment's project
ls_utils.get_tracer_project(),
)
run_create = {
**kwargs,
"session_name": project_name,
"name": name,
"inputs": inputs,
"run_type": run_type,
}
if not self._filter_for_sampling([run_create]):
return
if revision_id is not None:
run_create["extra"]["metadata"]["revision_id"] = revision_id
run_create = self._run_transform(
run_create,
copy=False,
)
self._insert_runtime_env([run_create])
if (
# batch ingest requires trace_id and dotted_order to be set
run_create.get("trace_id") is not None
and run_create.get("dotted_order") is not None
):
if self._pyo3_client is not None:
self._pyo3_client.create_run(run_create)
elif self.compressed_runs is not None:
serialized_op = serialize_run_dict("post", run_create)
multipart_form = (
serialized_run_operation_to_multipart_parts_and_context(
serialized_op
)
)
with self.compressed_runs.lock:
compress_multipart_parts_and_context(
multipart_form,
self.compressed_runs,
_BOUNDARY,
)
self.compressed_runs.run_count += 1
self._data_available_event.set()
elif self.tracing_queue is not None:
serialized_op = serialize_run_dict("post", run_create)
self.tracing_queue.put(
TracingQueueItem(run_create["dotted_order"], serialized_op)
)
else:
# Neither Rust nor Python batch ingestion is configured,
# fall back to the non-batch approach.
self._create_run(run_create)
else:
self._create_run(run_create)
def _create_run(self, run_create: dict):
for api_url, api_key in self._write_api_urls.items():
headers = {**self._headers, X_API_KEY: api_key}
self.request_with_retries(
"POST",
f"{api_url}/runs",
request_kwargs={
"data": _dumps_json(run_create),
"headers": headers,
},
to_ignore=(ls_utils.LangSmithConflictError,),
)
def _hide_run_inputs(self, inputs: dict):
if self._hide_inputs is True:
return {}
if self._anonymizer:
json_inputs = _orjson.loads(_dumps_json(inputs))
return self._anonymizer(json_inputs)
if self._hide_inputs is False:
return inputs
return self._hide_inputs(inputs)
def _hide_run_outputs(self, outputs: dict):
if self._hide_outputs is True:
return {}
if self._anonymizer:
json_outputs = _orjson.loads(_dumps_json(outputs))
return self._anonymizer(json_outputs)
if self._hide_outputs is False:
return outputs
return self._hide_outputs(outputs)
def _batch_ingest_run_ops(
self,
ops: List[SerializedRunOperation],
) -> None:
ids_and_partial_body: dict[
Literal["post", "patch"], list[tuple[str, bytes]]
] = {
"post": [],
"patch": [],
}
# form the partial body and ids
for op in ops:
if isinstance(op, SerializedRunOperation):
curr_dict = _orjson.loads(op._none)
if op.inputs:
curr_dict["inputs"] = _orjson.Fragment(op.inputs)
if op.outputs:
curr_dict["outputs"] = _orjson.Fragment(op.outputs)
if op.events:
curr_dict["events"] = _orjson.Fragment(op.events)
if op.attachments:
logger.warning(
"Attachments are not supported when use_multipart_endpoint "
"is False"
)
ids_and_partial_body[op.operation].append(
(f"trace={op.trace_id},id={op.id}", _orjson.dumps(curr_dict))
)
elif isinstance(op, SerializedFeedbackOperation):
logger.warning(
"Feedback operations are not supported in non-multipart mode"
)
else:
logger.error("Unknown item type in tracing queue: %s", type(op))
# send the requests in batches
info = self.info
size_limit_bytes = (info.batch_ingest_config or {}).get(
"size_limit_bytes"
) or _SIZE_LIMIT_BYTES
body_chunks: DefaultDict[str, list] = collections.defaultdict(list)
context_ids: DefaultDict[str, list] = collections.defaultdict(list)
body_size = 0
for key in cast(List[Literal["post", "patch"]], ["post", "patch"]):
body_deque = collections.deque(ids_and_partial_body[key])
while body_deque:
if (
body_size > 0
and body_size + len(body_deque[0][1]) > size_limit_bytes
):
self._post_batch_ingest_runs(
_orjson.dumps(body_chunks),
_context=f"\n{key}: {'; '.join(context_ids[key])}",
)
body_size = 0
body_chunks.clear()
context_ids.clear()
curr_id, curr_body = body_deque.popleft()
body_size += len(curr_body)
body_chunks[key].append(_orjson.Fragment(curr_body))
context_ids[key].append(curr_id)
if body_size:
context = "; ".join(f"{k}: {'; '.join(v)}" for k, v in context_ids.items())
self._post_batch_ingest_runs(
_orjson.dumps(body_chunks), _context="\n" + context
)
def batch_ingest_runs(
self,
create: Optional[
Sequence[Union[ls_schemas.Run, ls_schemas.RunLikeDict, Dict]]
] = None,
update: Optional[
Sequence[Union[ls_schemas.Run, ls_schemas.RunLikeDict, Dict]]
] = None,
*,
pre_sampled: bool = False,
) -> None:
"""Batch ingest/upsert multiple runs in the Langsmith system.
Args:
create (Optional[Sequence[Union[Run, RunLikeDict]]]):
A sequence of `Run` objects or equivalent dictionaries representing
runs to be created / posted.
update (Optional[Sequence[Union[Run, RunLikeDict]]]):
A sequence of `Run` objects or equivalent dictionaries representing
runs that have already been created and should be updated / patched.
pre_sampled (bool, default=False): Whether the runs have already been subject
to sampling, and therefore should not be sampled again.
Defaults to False.
Raises:
LangsmithAPIError: If there is an error in the API request.
Returns:
None
Note:
- The run objects MUST contain the dotted_order and trace_id fields
to be accepted by the API.
Examples:
.. code-block:: python
from langsmith import Client
import datetime
from uuid import uuid4
client = Client()
_session = "__test_batch_ingest_runs"
trace_id = uuid4()
trace_id_2 = uuid4()
run_id_2 = uuid4()
current_time = datetime.datetime.now(datetime.timezone.utc).strftime(
"%Y%m%dT%H%M%S%fZ"
)
later_time = (
datetime.datetime.now(datetime.timezone.utc) + timedelta(seconds=1)
).strftime("%Y%m%dT%H%M%S%fZ")
runs_to_create = [
{
"id": str(trace_id),
"session_name": _session,
"name": "run 1",
"run_type": "chain",
"dotted_order": f"{current_time}{str(trace_id)}",
"trace_id": str(trace_id),
"inputs": {"input1": 1, "input2": 2},
"outputs": {"output1": 3, "output2": 4},
},
{
"id": str(trace_id_2),
"session_name": _session,
"name": "run 3",
"run_type": "chain",
"dotted_order": f"{current_time}{str(trace_id_2)}",
"trace_id": str(trace_id_2),
"inputs": {"input1": 1, "input2": 2},
"error": "error",
},
{
"id": str(run_id_2),
"session_name": _session,
"name": "run 2",
"run_type": "chain",
"dotted_order": f"{current_time}{str(trace_id)}."
f"{later_time}{str(run_id_2)}",
"trace_id": str(trace_id),
"parent_run_id": str(trace_id),
"inputs": {"input1": 5, "input2": 6},
},
]
runs_to_update = [
{
"id": str(run_id_2),
"dotted_order": f"{current_time}{str(trace_id)}."
f"{later_time}{str(run_id_2)}",
"trace_id": str(trace_id),
"parent_run_id": str(trace_id),
"outputs": {"output1": 4, "output2": 5},
},
]
client.batch_ingest_runs(create=runs_to_create, update=runs_to_update)
"""
if not create and not update:
return
# transform and convert to dicts
create_dicts = [
self._run_transform(run, copy=False) for run in create or EMPTY_SEQ
]
update_dicts = [
self._run_transform(run, update=True, copy=False)
for run in update or EMPTY_SEQ
]
for run in create_dicts:
if not run.get("trace_id") or not run.get("dotted_order"):
raise ls_utils.LangSmithUserError(
"Batch ingest requires trace_id and dotted_order to be set."
)
for run in update_dicts:
if not run.get("trace_id") or not run.get("dotted_order"):
raise ls_utils.LangSmithUserError(
"Batch ingest requires trace_id and dotted_order to be set."
)
# filter out runs that are not sampled
if not pre_sampled:
create_dicts = self._filter_for_sampling(create_dicts)
update_dicts = self._filter_for_sampling(update_dicts, patch=True)
if not create_dicts and not update_dicts:
return
self._insert_runtime_env(create_dicts + update_dicts)
# convert to serialized ops
serialized_ops = cast(
List[SerializedRunOperation],
combine_serialized_queue_operations(
list(
itertools.chain(
(serialize_run_dict("post", run) for run in create_dicts),
(serialize_run_dict("patch", run) for run in update_dicts),
)
)
),
)
self._batch_ingest_run_ops(serialized_ops)
def _post_batch_ingest_runs(self, body: bytes, *, _context: str):
for api_url, api_key in self._write_api_urls.items():
try:
self.request_with_retries(
"POST",
f"{api_url}/runs/batch",
request_kwargs={
"data": body,
"headers": {
**self._headers,
X_API_KEY: api_key,
},
},
to_ignore=(ls_utils.LangSmithConflictError,),
stop_after_attempt=3,
_context=_context,
)
except Exception as e:
try:
exc_desc_lines = traceback.format_exception_only(type(e), e)
exc_desc = "".join(exc_desc_lines).rstrip()
logger.warning(f"Failed to batch ingest runs: {exc_desc}")
except Exception:
logger.warning(f"Failed to batch ingest runs: {repr(e)}")
def _multipart_ingest_ops(
self, ops: list[Union[SerializedRunOperation, SerializedFeedbackOperation]]
) -> None:
parts: list[MultipartPartsAndContext] = []
for op in ops:
if isinstance(op, SerializedRunOperation):
parts.append(
serialized_run_operation_to_multipart_parts_and_context(op)
)
elif isinstance(op, SerializedFeedbackOperation):
parts.append(
serialized_feedback_operation_to_multipart_parts_and_context(op)
)
else:
logger.error("Unknown operation type in tracing queue: %s", type(op))
acc_multipart = join_multipart_parts_and_context(parts)
if acc_multipart:
self._send_multipart_req(acc_multipart)
def multipart_ingest(
self,
create: Optional[
Sequence[Union[ls_schemas.Run, ls_schemas.RunLikeDict, Dict]]
] = None,
update: Optional[
Sequence[Union[ls_schemas.Run, ls_schemas.RunLikeDict, Dict]]
] = None,
*,
pre_sampled: bool = False,
) -> None:
"""Batch ingest/upsert multiple runs in the Langsmith system.
Args:
create (Optional[Sequence[Union[ls_schemas.Run, RunLikeDict]]]):
A sequence of `Run` objects or equivalent dictionaries representing
runs to be created / posted.
update (Optional[Sequence[Union[ls_schemas.Run, RunLikeDict]]]):
A sequence of `Run` objects or equivalent dictionaries representing
runs that have already been created and should be updated / patched.
pre_sampled (bool, default=False): Whether the runs have already been subject
to sampling, and therefore should not be sampled again.
Defaults to False.
Raises:
LangsmithAPIError: If there is an error in the API request.
Returns:
None
Note:
- The run objects MUST contain the dotted_order and trace_id fields
to be accepted by the API.
Examples:
.. code-block:: python
from langsmith import Client
import datetime
from uuid import uuid4
client = Client()
_session = "__test_batch_ingest_runs"
trace_id = uuid4()
trace_id_2 = uuid4()
run_id_2 = uuid4()
current_time = datetime.datetime.now(datetime.timezone.utc).strftime(
"%Y%m%dT%H%M%S%fZ"
)
later_time = (
datetime.datetime.now(datetime.timezone.utc) + timedelta(seconds=1)
).strftime("%Y%m%dT%H%M%S%fZ")
runs_to_create = [
{
"id": str(trace_id),
"session_name": _session,
"name": "run 1",
"run_type": "chain",
"dotted_order": f"{current_time}{str(trace_id)}",
"trace_id": str(trace_id),
"inputs": {"input1": 1, "input2": 2},
"outputs": {"output1": 3, "output2": 4},
},
{
"id": str(trace_id_2),
"session_name": _session,
"name": "run 3",
"run_type": "chain",
"dotted_order": f"{current_time}{str(trace_id_2)}",
"trace_id": str(trace_id_2),
"inputs": {"input1": 1, "input2": 2},
"error": "error",
},
{
"id": str(run_id_2),
"session_name": _session,
"name": "run 2",
"run_type": "chain",
"dotted_order": f"{current_time}{str(trace_id)}."
f"{later_time}{str(run_id_2)}",
"trace_id": str(trace_id),
"parent_run_id": str(trace_id),
"inputs": {"input1": 5, "input2": 6},
},
]
runs_to_update = [
{
"id": str(run_id_2),
"dotted_order": f"{current_time}{str(trace_id)}."
f"{later_time}{str(run_id_2)}",
"trace_id": str(trace_id),
"parent_run_id": str(trace_id),
"outputs": {"output1": 4, "output2": 5},
},
]
client.multipart_ingest(create=runs_to_create, update=runs_to_update)
"""
if not (create or update):
return
# transform and convert to dicts
create_dicts = [self._run_transform(run) for run in create or EMPTY_SEQ]
update_dicts = [
self._run_transform(run, update=True) for run in update or EMPTY_SEQ
]
# require trace_id and dotted_order
if create_dicts:
for run in create_dicts:
if not run.get("trace_id") or not run.get("dotted_order"):
raise ls_utils.LangSmithUserError(
"Multipart ingest requires trace_id and dotted_order"
" to be set in create dicts."
)
else:
del run
if update_dicts:
for run in update_dicts:
if not run.get("trace_id") or not run.get("dotted_order"):
raise ls_utils.LangSmithUserError(
"Multipart ingest requires trace_id and dotted_order"
" to be set in update dicts."
)
else:
del run
# combine post and patch dicts where possible
if update_dicts and create_dicts:
create_by_id = {run["id"]: run for run in create_dicts}
standalone_updates: list[dict] = []
for run in update_dicts:
if run["id"] in create_by_id:
for k, v in run.items():
if v is not None:
create_by_id[run["id"]][k] = v
else:
standalone_updates.append(run)
else:
del run
update_dicts = standalone_updates
# filter out runs that are not sampled
if not pre_sampled:
create_dicts = self._filter_for_sampling(create_dicts)
update_dicts = self._filter_for_sampling(update_dicts, patch=True)
if not create_dicts and not update_dicts:
return
# insert runtime environment
self._insert_runtime_env(create_dicts)
self._insert_runtime_env(update_dicts)
# format as serialized operations
serialized_ops = combine_serialized_queue_operations(
list(
itertools.chain(
(serialize_run_dict("post", run) for run in create_dicts),
(serialize_run_dict("patch", run) for run in update_dicts),
)
)
)
# sent the runs in multipart requests
self._multipart_ingest_ops(serialized_ops)
def _send_multipart_req(self, acc: MultipartPartsAndContext, *, attempts: int = 3):
parts = acc.parts
_context = acc.context
for api_url, api_key in self._write_api_urls.items():
for idx in range(1, attempts + 1):
try:
encoder = rqtb_multipart.MultipartEncoder(parts, boundary=_BOUNDARY)
if encoder.len <= 20_000_000: # ~20 MB
data = encoder.to_string()
else:
data = encoder
self.request_with_retries(
"POST",
f"{api_url}/runs/multipart",
request_kwargs={
"data": data,
"headers": {
**self._headers,
X_API_KEY: api_key,
"Content-Type": encoder.content_type,
},
},
stop_after_attempt=1,
_context=_context,
)
break
except ls_utils.LangSmithConflictError:
break
except (
ls_utils.LangSmithConnectionError,
ls_utils.LangSmithRequestTimeout,
ls_utils.LangSmithAPIError,
) as exc:
if idx == attempts:
logger.warning(f"Failed to multipart ingest runs: {exc}")
else:
continue
except Exception as e:
try:
exc_desc_lines = traceback.format_exception_only(type(e), e)
exc_desc = "".join(exc_desc_lines).rstrip()
logger.warning(f"Failed to multipart ingest runs: {exc_desc}")
except Exception:
logger.warning(f"Failed to multipart ingest runs: {repr(e)}")
# do not retry by default
return
def _send_compressed_multipart_req(
self,
data_stream: io.BytesIO,
compressed_runs_info: Optional[Tuple[int, int]],
*,
attempts: int = 3,
):
"""Send a zstd-compressed multipart form data stream to the backend."""
_context: str = ""
for api_url, api_key in self._write_api_urls.items():
data_stream.seek(0)
for idx in range(1, attempts + 1):
try:
headers = {
**self._headers,
"X-API-KEY": api_key,
"Content-Type": f"multipart/form-data; boundary={_BOUNDARY}",
"Content-Encoding": "zstd",
"X-Pre-Compressed-Size": (
str(compressed_runs_info[0]) if compressed_runs_info else ""
),
"X-Post-Compressed-Size": (
str(compressed_runs_info[1]) if compressed_runs_info else ""
),
}
self.request_with_retries(
"POST",
f"{api_url}/runs/multipart",
request_kwargs={
"data": data_stream,
"headers": headers,
},
stop_after_attempt=1,
_context=_context,
)
break
except ls_utils.LangSmithConflictError:
break
except (
ls_utils.LangSmithConnectionError,
ls_utils.LangSmithRequestTimeout,
ls_utils.LangSmithAPIError,
) as exc:
if idx == attempts:
logger.warning(
f"Failed to send compressed multipart ingest: {exc}"
)
else:
continue
except Exception as e:
try:
exc_desc_lines = traceback.format_exception_only(type(e), e)
exc_desc = "".join(exc_desc_lines).rstrip()
logger.warning(
f"Failed to send compressed multipart ingest: {exc_desc}"
)
except Exception:
logger.warning(
f"Failed to send compressed multipart ingest: {repr(e)}"
)
# Do not retry by default after unknown exceptions
return
def update_run(
self,
run_id: ID_TYPE,
*,
name: Optional[str] = None,
end_time: Optional[datetime.datetime] = None,
error: Optional[str] = None,
inputs: Optional[Dict] = None,
outputs: Optional[Dict] = None,
events: Optional[Sequence[dict]] = None,
extra: Optional[Dict] = None,
tags: Optional[List[str]] = None,
attachments: Optional[ls_schemas.Attachments] = None,
**kwargs: Any,
) -> None:
"""Update a run in the LangSmith API.
Args:
run_id (Union[UUID, str]): The ID of the run to update.
name (Optional[str]): The name of the run.
end_time (Optional[datetime.datetime]): The end time of the run.
error (Optional[str]): The error message of the run.
inputs (Optional[Dict]): The input values for the run.
outputs (Optional[Dict]): The output values for the run.
events (Optional[Sequence[dict]]): The events for the run.
extra (Optional[Dict]): The extra information for the run.
tags (Optional[List[str]]): The tags for the run.
attachments (Optional[Dict[str, Attachment]]): A dictionary of attachments to add to the run. The keys are the attachment names,
and the values are Attachment objects containing the data and mime type.
**kwargs (Any): Kwargs are ignored.
Returns:
None
Examples:
.. code-block:: python
from langsmith import Client
import datetime
from uuid import uuid4
client = Client()
project_name = "__test_update_run"
start_time = datetime.datetime.now()
revision_id = uuid4()
run: dict = dict(
id=uuid4(),
name="test_run",
run_type="llm",
inputs={"text": "hello world"},
project_name=project_name,
api_url=os.getenv("LANGCHAIN_ENDPOINT"),
start_time=start_time,
extra={"extra": "extra"},
revision_id=revision_id,
)
# Create the run
client.create_run(**run)
run["outputs"] = {"output": ["Hi"]}
run["extra"]["foo"] = "bar"
run["name"] = "test_run_updated"
# Update the run
client.update_run(run["id"], **run)
"""
data: Dict[str, Any] = {
"id": _as_uuid(run_id, "run_id"),
"name": name,
"trace_id": kwargs.pop("trace_id", None),
"parent_run_id": kwargs.pop("parent_run_id", None),
"dotted_order": kwargs.pop("dotted_order", None),
"tags": tags,
"extra": extra,
"session_id": kwargs.pop("session_id", None),
"session_name": kwargs.pop("session_name", None),
}
if attachments:
data["attachments"] = attachments
use_multipart = (
(self.tracing_queue is not None or self.compressed_runs is not None)
# batch ingest requires trace_id and dotted_order to be set
and data["trace_id"] is not None
and data["dotted_order"] is not None
)
if not self._filter_for_sampling([data], patch=True):
return
if end_time is not None:
data["end_time"] = end_time.isoformat()
else:
data["end_time"] = datetime.datetime.now(datetime.timezone.utc).isoformat()
if error is not None:
data["error"] = error
if inputs is not None:
data["inputs"] = self._hide_run_inputs(inputs)
if outputs is not None:
if not use_multipart:
outputs = ls_utils.deepish_copy(outputs)
data["outputs"] = self._hide_run_outputs(outputs)
if events is not None:
data["events"] = events
if data["extra"]:
self._insert_runtime_env([data])
if self._pyo3_client is not None:
self._pyo3_client.update_run(data)
elif use_multipart:
serialized_op = serialize_run_dict(operation="patch", payload=data)
if self.compressed_runs is not None:
multipart_form = (
serialized_run_operation_to_multipart_parts_and_context(
serialized_op
)
)
with self.compressed_runs.lock:
compress_multipart_parts_and_context(
multipart_form,
self.compressed_runs,
_BOUNDARY,
)
self.compressed_runs.run_count += 1
self._data_available_event.set()
elif self.tracing_queue is not None:
self.tracing_queue.put(
TracingQueueItem(data["dotted_order"], serialized_op)
)
else:
self._update_run(data)
def _update_run(self, run_update: dict) -> None:
for api_url, api_key in self._write_api_urls.items():
headers = {
**self._headers,
X_API_KEY: api_key,
}
self.request_with_retries(
"PATCH",
f"{api_url}/runs/{run_update['id']}",
request_kwargs={
"data": _dumps_json(run_update),
"headers": headers,
},
)
def flush_compressed_runs(self, attempts: int = 3) -> None:
"""Force flush the currently buffered compressed runs."""
if self.compressed_runs is None:
return
# Attempt to drain and send any remaining data
from langsmith._internal._background_thread import (
HTTP_REQUEST_THREAD_POOL,
_tracing_thread_drain_compressed_buffer,
)
final_data_stream, compressed_runs_info = (
_tracing_thread_drain_compressed_buffer(
self, size_limit=1, size_limit_bytes=1
)
)
if final_data_stream is not None:
# We have data to send
future = None
try:
future = HTTP_REQUEST_THREAD_POOL.submit(
self._send_compressed_multipart_req,
final_data_stream,
compressed_runs_info,
attempts=attempts,
)
self._futures.add(future)
except RuntimeError:
# In case the ThreadPoolExecutor is already shutdown
self._send_compressed_multipart_req(
final_data_stream, compressed_runs_info, attempts=attempts
)
# If we got a future, wait for it to complete
if self._futures:
done, _ = cf.wait(self._futures)
# Remove completed futures
self._futures.difference_update(done)
def flush(self) -> None:
"""Flush either queue or compressed buffer, depending on mode."""
if self.compressed_runs is not None:
self.flush_compressed_runs()
elif self.tracing_queue is not None:
self.tracing_queue.join()
def _load_child_runs(self, run: ls_schemas.Run) -> ls_schemas.Run:
"""Load child runs for a given run.
Args:
run (Run): The run to load child runs for.
Returns:
Run: The run with loaded child runs.
Raises:
LangSmithError: If a child run has no parent.
"""
child_runs = self.list_runs(id=run.child_run_ids)
treemap: DefaultDict[uuid.UUID, List[ls_schemas.Run]] = collections.defaultdict(
list
)
runs: Dict[uuid.UUID, ls_schemas.Run] = {}
for child_run in sorted(
child_runs,
key=lambda r: r.dotted_order,
):
if child_run.parent_run_id is None:
raise ls_utils.LangSmithError(f"Child run {child_run.id} has no parent")
treemap[child_run.parent_run_id].append(child_run)
runs[child_run.id] = child_run
run.child_runs = treemap.pop(run.id, [])
for run_id, children in treemap.items():
runs[run_id].child_runs = children
return run
def read_run(
self, run_id: ID_TYPE, load_child_runs: bool = False
) -> ls_schemas.Run:
"""Read a run from the LangSmith API.
Args:
run_id (Union[UUID, str]):
The ID of the run to read.
load_child_runs (bool, default=False):
Whether to load nested child runs.
Returns:
Run: The run read from the LangSmith API.
Examples:
.. code-block:: python
from langsmith import Client
# Existing run
run_id = "your-run-id"
client = Client()
stored_run = client.read_run(run_id)
"""
response = self.request_with_retries(
"GET", f"/runs/{_as_uuid(run_id, 'run_id')}"
)
run = ls_schemas.Run(**response.json(), _host_url=self._host_url)
if load_child_runs and run.child_run_ids:
run = self._load_child_runs(run)
return run
def list_runs(
self,
*,
project_id: Optional[Union[ID_TYPE, Sequence[ID_TYPE]]] = None,
project_name: Optional[Union[str, Sequence[str]]] = None,
run_type: Optional[str] = None,
trace_id: Optional[ID_TYPE] = None,
reference_example_id: Optional[ID_TYPE] = None,
query: Optional[str] = None,
filter: Optional[str] = None,
trace_filter: Optional[str] = None,
tree_filter: Optional[str] = None,
is_root: Optional[bool] = None,
parent_run_id: Optional[ID_TYPE] = None,
start_time: Optional[datetime.datetime] = None,
error: Optional[bool] = None,
run_ids: Optional[Sequence[ID_TYPE]] = None,
select: Optional[Sequence[str]] = None,
limit: Optional[int] = None,
**kwargs: Any,
) -> Iterator[ls_schemas.Run]:
"""List runs from the LangSmith API.
Args:
project_id (Optional[Union[UUID, str], Sequence[Union[UUID, str]]]):
The ID(s) of the project to filter by.
project_name (Optional[Union[str, Sequence[str]]]): The name(s) of the project to filter by.
run_type (Optional[str]): The type of the runs to filter by.
trace_id (Optional[Union[UUID, str]]): The ID of the trace to filter by.
reference_example_id (Optional[Union[UUID, str]]): The ID of the reference example to filter by.
query (Optional[str]): The query string to filter by.
filter (Optional[str]): The filter string to filter by.
trace_filter (Optional[str]): Filter to apply to the ROOT run in the trace tree. This is meant to
be used in conjunction with the regular `filter` parameter to let you
filter runs by attributes of the root run within a trace.
tree_filter (Optional[str]): Filter to apply to OTHER runs in the trace tree, including
sibling and child runs. This is meant to be used in conjunction with
the regular `filter` parameter to let you filter runs by attributes
of any run within a trace.
is_root (Optional[bool]): Whether to filter by root runs.
parent_run_id (Optional[Union[UUID, str]]):
The ID of the parent run to filter by.
start_time (Optional[datetime.datetime]):
The start time to filter by.
error (Optional[bool]): Whether to filter by error status.
run_ids (Optional[Sequence[Union[UUID, str]]]):
The IDs of the runs to filter by.
select (Optional[Sequence[str]]): The fields to select.
limit (Optional[int]): The maximum number of runs to return.
**kwargs (Any): Additional keyword arguments.
Yields:
The runs.
Examples:
.. code-block:: python
# List all runs in a project
project_runs = client.list_runs(project_name="")
# List LLM and Chat runs in the last 24 hours
todays_llm_runs = client.list_runs(
project_name="",
start_time=datetime.now() - timedelta(days=1),
run_type="llm",
)
# List root traces in a project
root_runs = client.list_runs(project_name="", is_root=1)
# List runs without errors
correct_runs = client.list_runs(project_name="", error=False)
# List runs and only return their inputs/outputs (to speed up the query)
input_output_runs = client.list_runs(
project_name="", select=["inputs", "outputs"]
)
# List runs by run ID
run_ids = [
"a36092d2-4ad5-4fb4-9c0d-0dba9a2ed836",
"9398e6be-964f-4aa4-8ae9-ad78cd4b7074",
]
selected_runs = client.list_runs(id=run_ids)
# List all "chain" type runs that took more than 10 seconds and had
# `total_tokens` greater than 5000
chain_runs = client.list_runs(
project_name="",
filter='and(eq(run_type, "chain"), gt(latency, 10), gt(total_tokens, 5000))',
)
# List all runs called "extractor" whose root of the trace was assigned feedback "user_score" score of 1
good_extractor_runs = client.list_runs(
project_name="",
filter='eq(name, "extractor")',
trace_filter='and(eq(feedback_key, "user_score"), eq(feedback_score, 1))',
)
# List all runs that started after a specific timestamp and either have "error" not equal to null or a "Correctness" feedback score equal to 0
complex_runs = client.list_runs(
project_name="",
filter='and(gt(start_time, "2023-07-15T12:34:56Z"), or(neq(error, null), and(eq(feedback_key, "Correctness"), eq(feedback_score, 0.0))))',
)
# List all runs where `tags` include "experimental" or "beta" and `latency` is greater than 2 seconds
tagged_runs = client.list_runs(
project_name="",
filter='and(or(has(tags, "experimental"), has(tags, "beta")), gt(latency, 2))',
)
""" # noqa: E501
project_ids = []
if isinstance(project_id, (uuid.UUID, str)):
project_ids.append(project_id)
elif isinstance(project_id, list):
project_ids.extend(project_id)
if project_name is not None:
if isinstance(project_name, str):
project_name = [project_name]
project_ids.extend(
[self.read_project(project_name=name).id for name in project_name]
)
default_select = [
"app_path",
"child_run_ids",
"completion_cost",
"completion_tokens",
"dotted_order",
"end_time",
"error",
"events",
"extra",
"feedback_stats",
"first_token_time",
"id",
"inputs",
"name",
"outputs",
"parent_run_id",
"parent_run_ids",
"prompt_cost",
"prompt_tokens",
"reference_example_id",
"run_type",
"session_id",
"start_time",
"status",
"tags",
"total_cost",
"total_tokens",
"trace_id",
]
select = select or default_select
body_query: Dict[str, Any] = {
"session": project_ids if project_ids else None,
"run_type": run_type,
"reference_example": (
[reference_example_id] if reference_example_id else None
),
"query": query,
"filter": filter,
"trace_filter": trace_filter,
"tree_filter": tree_filter,
"is_root": is_root,
"parent_run": parent_run_id,
"start_time": start_time.isoformat() if start_time else None,
"error": error,
"id": run_ids,
"trace": trace_id,
"select": select,
**kwargs,
}
body_query = {k: v for k, v in body_query.items() if v is not None}
for i, run in enumerate(
self._get_cursor_paginated_list("/runs/query", body=body_query)
):
yield ls_schemas.Run(**run, _host_url=self._host_url)
if limit is not None and i + 1 >= limit:
break
def get_run_stats(
self,
*,
id: Optional[List[ID_TYPE]] = None,
trace: Optional[ID_TYPE] = None,
parent_run: Optional[ID_TYPE] = None,
run_type: Optional[str] = None,
project_names: Optional[List[str]] = None,
project_ids: Optional[List[ID_TYPE]] = None,
reference_example_ids: Optional[List[ID_TYPE]] = None,
start_time: Optional[str] = None,
end_time: Optional[str] = None,
error: Optional[bool] = None,
query: Optional[str] = None,
filter: Optional[str] = None,
trace_filter: Optional[str] = None,
tree_filter: Optional[str] = None,
is_root: Optional[bool] = None,
data_source_type: Optional[str] = None,
) -> Dict[str, Any]:
"""Get aggregate statistics over queried runs.
Takes in similar query parameters to `list_runs` and returns statistics
based on the runs that match the query.
Args:
id (Optional[List[Union[UUID, str]]]): List of run IDs to filter by.
trace (Optional[Union[UUID, str]]): Trace ID to filter by.
parent_run (Optional[Union[UUID, str]]): Parent run ID to filter by.
run_type (Optional[str]): Run type to filter by.
project_names (Optional[List[str]]): List of project names to filter by.
project_ids (Optional[List[Union[UUID, str]]]): List of project IDs to filter by.
reference_example_ids (Optional[List[Union[UUID, str]]]): List of reference example IDs to filter by.
start_time (Optional[str]): Start time to filter by.
end_time (Optional[str]): End time to filter by.
error (Optional[bool]): Filter by error status.
query (Optional[str]): Query string to filter by.
filter (Optional[str]): Filter string to apply.
trace_filter (Optional[str]): Trace filter string to apply.
tree_filter (Optional[str]): Tree filter string to apply.
is_root (Optional[bool]): Filter by root run status.
data_source_type (Optional[str]): Data source type to filter by.
Returns:
Dict[str, Any]: A dictionary containing the run statistics.
""" # noqa: E501
from concurrent.futures import ThreadPoolExecutor, as_completed # type: ignore
project_ids = project_ids or []
if project_names:
with ThreadPoolExecutor() as executor:
futures = [
executor.submit(self.read_project, project_name=name)
for name in project_names
]
for future in as_completed(futures):
project_ids.append(future.result().id)
payload = {
"id": id,
"trace": trace,
"parent_run": parent_run,
"run_type": run_type,
"session": project_ids,
"reference_example": reference_example_ids,
"start_time": start_time,
"end_time": end_time,
"error": error,
"query": query,
"filter": filter,
"trace_filter": trace_filter,
"tree_filter": tree_filter,
"is_root": is_root,
"data_source_type": data_source_type,
}
# Remove None values from the payload
payload = {k: v for k, v in payload.items() if v is not None}
response = self.request_with_retries(
"POST",
"/runs/stats",
request_kwargs={
"data": _dumps_json(payload),
},
)
ls_utils.raise_for_status_with_text(response)
return response.json()
def get_run_url(
self,
*,
run: ls_schemas.RunBase,
project_name: Optional[str] = None,
project_id: Optional[ID_TYPE] = None,
) -> str:
"""Get the URL for a run.
Not recommended for use within your agent runtime.
More for use interacting with runs after the fact
for data analysis or ETL workloads.
Args:
run (RunBase): The run.
project_name (Optional[str]): The name of the project.
project_id (Optional[Union[UUID, str]]): The ID of the project.
Returns:
str: The URL for the run.
"""
if session_id := getattr(run, "session_id", None):
pass
elif session_name := getattr(run, "session_name", None):
session_id = self.read_project(project_name=session_name).id
elif project_id is not None:
session_id = project_id
elif project_name is not None:
session_id = self.read_project(project_name=project_name).id
else:
project_name = ls_utils.get_tracer_project()
session_id = self.read_project(project_name=project_name).id
session_id_ = _as_uuid(session_id, "session_id")
return (
f"{self._host_url}/o/{self._get_tenant_id()}/projects/p/{session_id_}/"
f"r/{run.id}?poll=true"
)
def share_run(self, run_id: ID_TYPE, *, share_id: Optional[ID_TYPE] = None) -> str:
"""Get a share link for a run.
Args:
run_id (Union[UUID, str]): The ID of the run to share.
share_id (Optional[Union[UUID, str]]): Custom share ID.
If not provided, a random UUID will be generated.
Returns:
str: The URL of the shared run.
"""
run_id_ = _as_uuid(run_id, "run_id")
data = {
"run_id": str(run_id_),
"share_token": share_id or str(uuid.uuid4()),
}
response = self.request_with_retries(
"PUT",
f"/runs/{run_id_}/share",
headers=self._headers,
json=data,
)
ls_utils.raise_for_status_with_text(response)
share_token = response.json()["share_token"]
return f"{self._host_url}/public/{share_token}/r"
def unshare_run(self, run_id: ID_TYPE) -> None:
"""Delete share link for a run.
Args:
run_id (Union[UUID, str]): The ID of the run to unshare.
Returns:
None
"""
response = self.request_with_retries(
"DELETE",
f"/runs/{_as_uuid(run_id, 'run_id')}/share",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
def read_run_shared_link(self, run_id: ID_TYPE) -> Optional[str]:
"""Retrieve the shared link for a specific run.
Args:
run_id (Union[UUID, str]): The ID of the run.
Returns:
Optional[str]: The shared link for the run, or None if the link is not
available.
"""
response = self.request_with_retries(
"GET",
f"/runs/{_as_uuid(run_id, 'run_id')}/share",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
result = response.json()
if result is None or "share_token" not in result:
return None
return f"{self._host_url}/public/{result['share_token']}/r"
def run_is_shared(self, run_id: ID_TYPE) -> bool:
"""Get share state for a run.
Args:
run_id (Union[UUID, str]): The ID of the run.
Returns:
bool: True if the run is shared, False otherwise.
"""
link = self.read_run_shared_link(_as_uuid(run_id, "run_id"))
return link is not None
def read_shared_run(
self, share_token: Union[ID_TYPE, str], run_id: Optional[ID_TYPE] = None
) -> ls_schemas.Run:
"""Get shared runs.
Args:
share_token (Union[UUID, str]): The share token or URL of the shared run.
run_id (Optional[Union[UUID, str]]): The ID of the specific run to retrieve.
If not provided, the full shared run will be returned.
Returns:
Run: The shared run.
"""
_, token_uuid = _parse_token_or_url(share_token, "", kind="run")
path = f"/public/{token_uuid}/run"
if run_id is not None:
path += f"/{_as_uuid(run_id, 'run_id')}"
response = self.request_with_retries(
"GET",
path,
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.Run(**response.json(), _host_url=self._host_url)
def list_shared_runs(
self, share_token: Union[ID_TYPE, str], run_ids: Optional[List[str]] = None
) -> Iterator[ls_schemas.Run]:
"""Get shared runs.
Args:
share_token (Union[UUID, str]): The share token or URL of the shared run.
run_ids (Optional[List[str]]): A list of run IDs to filter the results by.
Yields:
A shared run.
"""
body = {"id": run_ids} if run_ids else {}
_, token_uuid = _parse_token_or_url(share_token, "", kind="run")
for run in self._get_cursor_paginated_list(
f"/public/{token_uuid}/runs/query", body=body
):
yield ls_schemas.Run(**run, _host_url=self._host_url)
def read_dataset_shared_schema(
self,
dataset_id: Optional[ID_TYPE] = None,
*,
dataset_name: Optional[str] = None,
) -> ls_schemas.DatasetShareSchema:
"""Retrieve the shared schema of a dataset.
Args:
dataset_id (Optional[Union[UUID, str]]): The ID of the dataset.
Either `dataset_id` or `dataset_name` must be given.
dataset_name (Optional[str]): The name of the dataset.
Either `dataset_id` or `dataset_name` must be given.
Returns:
ls_schemas.DatasetShareSchema: The shared schema of the dataset.
Raises:
ValueError: If neither `dataset_id` nor `dataset_name` is given.
"""
if dataset_id is None and dataset_name is None:
raise ValueError("Either dataset_id or dataset_name must be given")
if dataset_id is None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
response = self.request_with_retries(
"GET",
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/share",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
d = response.json()
return cast(
ls_schemas.DatasetShareSchema,
{
**d,
"url": f"{self._host_url}/public/"
f"{_as_uuid(d['share_token'], 'response.share_token')}/d",
},
)
def share_dataset(
self,
dataset_id: Optional[ID_TYPE] = None,
*,
dataset_name: Optional[str] = None,
) -> ls_schemas.DatasetShareSchema:
"""Get a share link for a dataset.
Args:
dataset_id (Optional[Union[UUID, str]]): The ID of the dataset.
Either `dataset_id` or `dataset_name` must be given.
dataset_name (Optional[str]): The name of the dataset.
Either `dataset_id` or `dataset_name` must be given.
Returns:
ls_schemas.DatasetShareSchema: The shared schema of the dataset.
Raises:
ValueError: If neither `dataset_id` nor `dataset_name` is given.
"""
if dataset_id is None and dataset_name is None:
raise ValueError("Either dataset_id or dataset_name must be given")
if dataset_id is None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
data = {
"dataset_id": str(dataset_id),
}
response = self.request_with_retries(
"PUT",
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/share",
headers=self._headers,
json=data,
)
ls_utils.raise_for_status_with_text(response)
d: dict = response.json()
return cast(
ls_schemas.DatasetShareSchema,
{**d, "url": f"{self._host_url}/public/{d['share_token']}/d"},
)
def unshare_dataset(self, dataset_id: ID_TYPE) -> None:
"""Delete share link for a dataset.
Args:
dataset_id (Union[UUID, str]): The ID of the dataset to unshare.
Returns:
None
"""
response = self.request_with_retries(
"DELETE",
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/share",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
def read_shared_dataset(
self,
share_token: str,
) -> ls_schemas.Dataset:
"""Get shared datasets.
Args:
share_token (Union[UUID, str]): The share token or URL of the shared dataset.
Returns:
Dataset: The shared dataset.
"""
_, token_uuid = _parse_token_or_url(share_token, self.api_url)
response = self.request_with_retries(
"GET",
f"/public/{token_uuid}/datasets",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.Dataset(
**response.json(),
_host_url=self._host_url,
_public_path=f"/public/{share_token}/d",
)
def list_shared_examples(
self, share_token: str, *, example_ids: Optional[List[ID_TYPE]] = None
) -> List[ls_schemas.Example]:
"""Get shared examples.
Args:
share_token (Union[UUID, str]): The share token or URL of the shared dataset.
example_ids (Optional[List[UUID, str]], optional): The IDs of the examples to filter by. Defaults to None.
Returns:
List[ls_schemas.Example]: The list of shared examples.
"""
params = {}
if example_ids is not None:
params["id"] = [str(id) for id in example_ids]
response = self.request_with_retries(
"GET",
f"/public/{_as_uuid(share_token, 'share_token')}/examples",
headers=self._headers,
params=params,
)
ls_utils.raise_for_status_with_text(response)
return [
ls_schemas.Example(**dataset, _host_url=self._host_url)
for dataset in response.json()
]
def list_shared_projects(
self,
*,
dataset_share_token: str,
project_ids: Optional[List[ID_TYPE]] = None,
name: Optional[str] = None,
name_contains: Optional[str] = None,
limit: Optional[int] = None,
) -> Iterator[ls_schemas.TracerSessionResult]:
"""List shared projects.
Args:
dataset_share_token (str): The share token of the dataset.
project_ids (Optional[List[Union[UUID, str]]]): List of project IDs to filter the results, by default None.
name (Optional[str]): Name of the project to filter the results, by default None.
name_contains (Optional[str]): Substring to search for in project names, by default None.
limit (Optional[int]): Maximum number of projects to return, by default None.
Yields:
The shared projects.
"""
params = {"id": project_ids, "name": name, "name_contains": name_contains}
share_token = _as_uuid(dataset_share_token, "dataset_share_token")
for i, project in enumerate(
self._get_paginated_list(
f"/public/{share_token}/datasets/sessions",
params=params,
)
):
yield ls_schemas.TracerSessionResult(**project, _host_url=self._host_url)
if limit is not None and i + 1 >= limit:
break
def create_project(
self,
project_name: str,
*,
description: Optional[str] = None,
metadata: Optional[dict] = None,
upsert: bool = False,
project_extra: Optional[dict] = None,
reference_dataset_id: Optional[ID_TYPE] = None,
) -> ls_schemas.TracerSession:
"""Create a project on the LangSmith API.
Args:
project_name (str): The name of the project.
project_extra (Optional[dict]): Additional project information.
metadata (Optional[dict]): Additional metadata to associate with the project.
description (Optional[str]): The description of the project.
upsert (bool, default=False): Whether to update the project if it already exists.
reference_dataset_id (Optional[Union[UUID, str]): The ID of the reference dataset to associate with the project.
Returns:
TracerSession: The created project.
"""
endpoint = f"{self.api_url}/sessions"
extra = project_extra
if metadata:
extra = {**(extra or {}), "metadata": metadata}
body: Dict[str, Any] = {
"name": project_name,
"extra": extra,
"description": description,
"id": str(uuid.uuid4()),
}
params = {}
if upsert:
params["upsert"] = True
if reference_dataset_id is not None:
body["reference_dataset_id"] = reference_dataset_id
response = self.request_with_retries(
"POST",
endpoint,
headers={**self._headers, "Content-Type": "application/json"},
data=_dumps_json(body),
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.TracerSession(**response.json(), _host_url=self._host_url)
def update_project(
self,
project_id: ID_TYPE,
*,
name: Optional[str] = None,
description: Optional[str] = None,
metadata: Optional[dict] = None,
project_extra: Optional[dict] = None,
end_time: Optional[datetime.datetime] = None,
) -> ls_schemas.TracerSession:
"""Update a LangSmith project.
Args:
project_id (Union[UUID, str]):
The ID of the project to update.
name (Optional[str]):
The new name to give the project. This is only valid if the project
has been assigned an end_time, meaning it has been completed/closed.
description (Optional[str]):
The new description to give the project.
metadata (Optional[dict]):
Additional metadata to associate with the project.
project_extra (Optional[dict]):
Additional project information.
end_time (Optional[datetime.datetime]):
The time the project was completed.
Returns:
TracerSession: The updated project.
"""
endpoint = f"{self.api_url}/sessions/{_as_uuid(project_id, 'project_id')}"
extra = project_extra
if metadata:
extra = {**(extra or {}), "metadata": metadata}
body: Dict[str, Any] = {
"name": name,
"extra": extra,
"description": description,
"end_time": end_time.isoformat() if end_time else None,
}
response = self.request_with_retries(
"PATCH",
endpoint,
headers={**self._headers, "Content-Type": "application/json"},
data=_dumps_json(body),
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.TracerSession(**response.json(), _host_url=self._host_url)
def _get_optional_tenant_id(self) -> Optional[uuid.UUID]:
if self._tenant_id is not None:
return self._tenant_id
try:
response = self.request_with_retries(
"GET", "/sessions", params={"limit": 1}
)
result = response.json()
if isinstance(result, list) and len(result) > 0:
tracer_session = ls_schemas.TracerSessionResult(
**result[0], _host_url=self._host_url
)
self._tenant_id = tracer_session.tenant_id
return self._tenant_id
except Exception as e:
logger.debug(
"Failed to get tenant ID from LangSmith: %s", repr(e), exc_info=True
)
return None
def _get_tenant_id(self) -> uuid.UUID:
tenant_id = self._get_optional_tenant_id()
if tenant_id is None:
raise ls_utils.LangSmithError("No tenant ID found")
return tenant_id
@ls_utils.xor_args(("project_id", "project_name"))
def read_project(
self,
*,
project_id: Optional[str] = None,
project_name: Optional[str] = None,
include_stats: bool = False,
) -> ls_schemas.TracerSessionResult:
"""Read a project from the LangSmith API.
Args:
project_id (Optional[str]):
The ID of the project to read.
project_name (Optional[str]): The name of the project to read.
Only one of project_id or project_name may be given.
include_stats (bool, default=False):
Whether to include a project's aggregate statistics in the response.
Returns:
TracerSessionResult: The project.
"""
path = "/sessions"
params: Dict[str, Any] = {"limit": 1}
if project_id is not None:
path += f"/{_as_uuid(project_id, 'project_id')}"
elif project_name is not None:
params["name"] = project_name
else:
raise ValueError("Must provide project_name or project_id")
params["include_stats"] = include_stats
response = self.request_with_retries("GET", path, params=params)
result = response.json()
if isinstance(result, list):
if len(result) == 0:
raise ls_utils.LangSmithNotFoundError(
f"Project {project_name} not found"
)
return ls_schemas.TracerSessionResult(**result[0], _host_url=self._host_url)
return ls_schemas.TracerSessionResult(
**response.json(), _host_url=self._host_url
)
def has_project(
self, project_name: str, *, project_id: Optional[str] = None
) -> bool:
"""Check if a project exists.
Args:
project_name (str):
The name of the project to check for.
project_id (Optional[str]):
The ID of the project to check for.
Returns:
bool: Whether the project exists.
"""
try:
self.read_project(project_name=project_name)
except ls_utils.LangSmithNotFoundError:
return False
return True
def get_test_results(
self,
*,
project_id: Optional[ID_TYPE] = None,
project_name: Optional[str] = None,
) -> pd.DataFrame:
"""Read the record-level information from an experiment into a Pandas DF.
Note: this will fetch whatever data exists in the DB. Results are not
immediately available in the DB upon evaluation run completion.
Args:
project_id (Optional[Union[UUID, str]]): The ID of the project.
project_name (Optional[str]): The name of the project.
Returns:
pd.DataFrame: A dataframe containing the test results.
"""
warnings.warn(
"Function get_test_results is in beta.", UserWarning, stacklevel=2
)
from concurrent.futures import ThreadPoolExecutor, as_completed # type: ignore
import pandas as pd # type: ignore
runs = self.list_runs(
project_id=project_id,
project_name=project_name,
is_root=True,
select=[
"id",
"reference_example_id",
"inputs",
"outputs",
"error",
"feedback_stats",
"start_time",
"end_time",
],
)
results: list[dict] = []
example_ids = []
def fetch_examples(batch):
examples = self.list_examples(example_ids=batch)
return [
{
"example_id": example.id,
**{f"reference.{k}": v for k, v in (example.outputs or {}).items()},
}
for example in examples
]
batch_size = 50
cursor = 0
with ThreadPoolExecutor() as executor:
futures = []
for r in runs:
row = {
"example_id": r.reference_example_id,
**{f"input.{k}": v for k, v in r.inputs.items()},
**{f"outputs.{k}": v for k, v in (r.outputs or {}).items()},
"execution_time": (
(r.end_time - r.start_time).total_seconds()
if r.end_time
else None
),
"error": r.error,
"id": r.id,
}
if r.feedback_stats:
row.update(
{
f"feedback.{k}": v.get("avg")
for k, v in r.feedback_stats.items()
}
)
if r.reference_example_id:
example_ids.append(r.reference_example_id)
else:
logger.warning(f"Run {r.id} has no reference example ID.")
if len(example_ids) % batch_size == 0:
# Ensure not empty
if batch := example_ids[cursor : cursor + batch_size]:
futures.append(executor.submit(fetch_examples, batch))
cursor += batch_size
results.append(row)
# Handle any remaining examples
if example_ids[cursor:]:
futures.append(executor.submit(fetch_examples, example_ids[cursor:]))
result_df = pd.DataFrame(results).set_index("example_id")
example_outputs = [
output for future in as_completed(futures) for output in future.result()
]
if example_outputs:
example_df = pd.DataFrame(example_outputs).set_index("example_id")
result_df = example_df.merge(result_df, left_index=True, right_index=True)
# Flatten dict columns into dot syntax for easier access
return pd.json_normalize(result_df.to_dict(orient="records"))
def list_projects(
self,
project_ids: Optional[List[ID_TYPE]] = None,
name: Optional[str] = None,
name_contains: Optional[str] = None,
reference_dataset_id: Optional[ID_TYPE] = None,
reference_dataset_name: Optional[str] = None,
reference_free: Optional[bool] = None,
limit: Optional[int] = None,
metadata: Optional[Dict[str, Any]] = None,
) -> Iterator[ls_schemas.TracerSession]:
"""List projects from the LangSmith API.
Args:
project_ids (Optional[List[Union[UUID, str]]]):
A list of project IDs to filter by, by default None
name (Optional[str]):
The name of the project to filter by, by default None
name_contains (Optional[str]):
A string to search for in the project name, by default None
reference_dataset_id (Optional[List[Union[UUID, str]]]):
A dataset ID to filter by, by default None
reference_dataset_name (Optional[str]):
The name of the reference dataset to filter by, by default None
reference_free (Optional[bool]):
Whether to filter for only projects not associated with a dataset.
limit (Optional[int]):
The maximum number of projects to return, by default None
metadata (Optional[Dict[str, Any]]):
Metadata to filter by.
Yields:
The projects.
Raises:
ValueError: If both reference_dataset_id and reference_dataset_name are given.
"""
params: Dict[str, Any] = {
"limit": min(limit, 100) if limit is not None else 100
}
if project_ids is not None:
params["id"] = project_ids
if name is not None:
params["name"] = name
if name_contains is not None:
params["name_contains"] = name_contains
if reference_dataset_id is not None:
if reference_dataset_name is not None:
raise ValueError(
"Only one of reference_dataset_id or"
" reference_dataset_name may be given"
)
params["reference_dataset"] = reference_dataset_id
elif reference_dataset_name is not None:
reference_dataset_id = self.read_dataset(
dataset_name=reference_dataset_name
).id
params["reference_dataset"] = reference_dataset_id
if reference_free is not None:
params["reference_free"] = reference_free
if metadata is not None:
params["metadata"] = json.dumps(metadata)
for i, project in enumerate(
self._get_paginated_list("/sessions", params=params)
):
yield ls_schemas.TracerSession(**project, _host_url=self._host_url)
if limit is not None and i + 1 >= limit:
break
@ls_utils.xor_args(("project_name", "project_id"))
def delete_project(
self, *, project_name: Optional[str] = None, project_id: Optional[str] = None
) -> None:
"""Delete a project from LangSmith.
Args:
project_name (Optional[str]):
The name of the project to delete.
project_id (Optional[str]):
The ID of the project to delete.
Returns:
None
Raises:
ValueError: If neither project_name or project_id is provided.
"""
if project_name is not None:
project_id = str(self.read_project(project_name=project_name).id)
elif project_id is None:
raise ValueError("Must provide project_name or project_id")
response = self.request_with_retries(
"DELETE",
f"/sessions/{_as_uuid(project_id, 'project_id')}",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
def create_dataset(
self,
dataset_name: str,
*,
description: Optional[str] = None,
data_type: ls_schemas.DataType = ls_schemas.DataType.kv,
inputs_schema: Optional[Dict[str, Any]] = None,
outputs_schema: Optional[Dict[str, Any]] = None,
transformations: Optional[List[ls_schemas.DatasetTransformation]] = None,
metadata: Optional[dict] = None,
) -> ls_schemas.Dataset:
"""Create a dataset in the LangSmith API.
Args:
dataset_name (str):
The name of the dataset.
description (Optional[str]):
The description of the dataset.
data_type (DataType, default=DataType.kv):
The data type of the dataset.
inputs_schema (Optional[Dict[str, Any]]):
The schema definition for the inputs of the dataset.
outputs_schema (Optional[Dict[str, Any]]):
The schema definition for the outputs of the dataset.
transformations (Optional[List[DatasetTransformation]]):
A list of transformations to apply to the dataset.
metadata (Optional[dict]):
Additional metadata to associate with the dataset.
Returns:
Dataset: The created dataset.
Raises:
requests.HTTPError: If the request to create the dataset fails.
"""
dataset: Dict[str, Any] = {
"name": dataset_name,
"data_type": data_type.value,
"created_at": datetime.datetime.now().isoformat(),
"transformations": transformations,
"extra": {"metadata": metadata} if metadata else None,
}
if description is not None:
dataset["description"] = description
if inputs_schema is not None:
dataset["inputs_schema_definition"] = inputs_schema
if outputs_schema is not None:
dataset["outputs_schema_definition"] = outputs_schema
response = self.request_with_retries(
"POST",
"/datasets",
headers={**self._headers, "Content-Type": "application/json"},
data=_orjson.dumps(dataset),
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.Dataset(
**response.json(),
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
def has_dataset(
self,
*,
dataset_name: Optional[str] = None,
dataset_id: Optional[ID_TYPE] = None,
) -> bool:
"""Check whether a dataset exists in your tenant.
Args:
dataset_name (Optional[str]):
The name of the dataset to check.
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to check.
Returns:
bool: Whether the dataset exists.
"""
try:
self.read_dataset(dataset_name=dataset_name, dataset_id=dataset_id)
return True
except ls_utils.LangSmithNotFoundError:
return False
@ls_utils.xor_args(("dataset_name", "dataset_id"))
def read_dataset(
self,
*,
dataset_name: Optional[str] = None,
dataset_id: Optional[ID_TYPE] = None,
) -> ls_schemas.Dataset:
"""Read a dataset from the LangSmith API.
Args:
dataset_name (Optional[str]):
The name of the dataset to read.
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to read.
Returns:
Dataset: The dataset.
"""
path = "/datasets"
params: Dict[str, Any] = {"limit": 1}
if dataset_id is not None:
path += f"/{_as_uuid(dataset_id, 'dataset_id')}"
elif dataset_name is not None:
params["name"] = dataset_name
else:
raise ValueError("Must provide dataset_name or dataset_id")
response = self.request_with_retries(
"GET",
path,
params=params,
)
result = response.json()
if isinstance(result, list):
if len(result) == 0:
raise ls_utils.LangSmithNotFoundError(
f"Dataset {dataset_name} not found"
)
return ls_schemas.Dataset(
**result[0],
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
return ls_schemas.Dataset(
**result,
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
def diff_dataset_versions(
self,
dataset_id: Optional[ID_TYPE] = None,
*,
dataset_name: Optional[str] = None,
from_version: Union[str, datetime.datetime],
to_version: Union[str, datetime.datetime],
) -> ls_schemas.DatasetDiffInfo:
"""Get the difference between two versions of a dataset.
Args:
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset.
dataset_name (Optional[str]):
The name of the dataset.
from_version (Union[str, datetime.datetime]):
The starting version for the diff.
to_version (Union[str, datetime.datetime]):
The ending version for the diff.
Returns:
DatasetDiffInfo: The difference between the two versions of the dataset.
Examples:
.. code-block:: python
# Get the difference between two tagged versions of a dataset
from_version = "prod"
to_version = "dev"
diff = client.diff_dataset_versions(
dataset_name="my-dataset",
from_version=from_version,
to_version=to_version,
)
# Get the difference between two timestamped versions of a dataset
from_version = datetime.datetime(2024, 1, 1)
to_version = datetime.datetime(2024, 2, 1)
diff = client.diff_dataset_versions(
dataset_name="my-dataset",
from_version=from_version,
to_version=to_version,
)
"""
if dataset_id is None:
if dataset_name is None:
raise ValueError("Must provide either dataset name or ID")
dataset_id = self.read_dataset(dataset_name=dataset_name).id
dsid = _as_uuid(dataset_id, "dataset_id")
response = self.request_with_retries(
"GET",
f"/datasets/{dsid}/versions/diff",
headers=self._headers,
params={
"from_version": (
from_version.isoformat()
if isinstance(from_version, datetime.datetime)
else from_version
),
"to_version": (
to_version.isoformat()
if isinstance(to_version, datetime.datetime)
else to_version
),
},
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.DatasetDiffInfo(**response.json())
def read_dataset_openai_finetuning(
self,
dataset_id: Optional[ID_TYPE] = None,
*,
dataset_name: Optional[str] = None,
) -> list:
"""Download a dataset in OpenAI Jsonl format and load it as a list of dicts.
Args:
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to download.
dataset_name (Optional[str]):
The name of the dataset to download.
Returns:
list[dict]: The dataset loaded as a list of dicts.
Raises:
ValueError: If neither dataset_id nor dataset_name is provided.
"""
path = "/datasets"
if dataset_id is not None:
pass
elif dataset_name is not None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
else:
raise ValueError("Must provide dataset_name or dataset_id")
response = self.request_with_retries(
"GET",
f"{path}/{_as_uuid(dataset_id, 'dataset_id')}/openai_ft",
)
dataset = [json.loads(line) for line in response.text.strip().split("\n")]
return dataset
def list_datasets(
self,
*,
dataset_ids: Optional[List[ID_TYPE]] = None,
data_type: Optional[str] = None,
dataset_name: Optional[str] = None,
dataset_name_contains: Optional[str] = None,
metadata: Optional[Dict[str, Any]] = None,
limit: Optional[int] = None,
) -> Iterator[ls_schemas.Dataset]:
"""List the datasets on the LangSmith API.
Args:
dataset_ids (Optional[List[Union[UUID, str]]]):
A list of dataset IDs to filter the results by.
data_type (Optional[str]):
The data type of the datasets to filter the results by.
dataset_name (Optional[str]):
The name of the dataset to filter the results by.
dataset_name_contains (Optional[str]):
A substring to search for in the dataset names.
metadata (Optional[Dict[str, Any]]):
A dictionary of metadata to filter the results by.
limit (Optional[int]):
The maximum number of datasets to return.
Yields:
The datasets.
"""
params: Dict[str, Any] = {
"limit": min(limit, 100) if limit is not None else 100
}
if dataset_ids is not None:
params["id"] = dataset_ids
if data_type is not None:
params["data_type"] = data_type
if dataset_name is not None:
params["name"] = dataset_name
if dataset_name_contains is not None:
params["name_contains"] = dataset_name_contains
if metadata is not None:
params["metadata"] = json.dumps(metadata)
for i, dataset in enumerate(
self._get_paginated_list("/datasets", params=params)
):
yield ls_schemas.Dataset(
**dataset,
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
if limit is not None and i + 1 >= limit:
break
@ls_utils.xor_args(("dataset_id", "dataset_name"))
def delete_dataset(
self,
*,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
) -> None:
"""Delete a dataset from the LangSmith API.
Args:
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to delete.
dataset_name (Optional[str]):
The name of the dataset to delete.
Returns:
None
"""
if dataset_name is not None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
if dataset_id is None:
raise ValueError("Must provide either dataset name or ID")
response = self.request_with_retries(
"DELETE",
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
def update_dataset_tag(
self,
*,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
as_of: datetime.datetime,
tag: str,
) -> None:
"""Update the tags of a dataset.
If the tag is already assigned to a different version of this dataset,
the tag will be moved to the new version. The as_of parameter is used to
determine which version of the dataset to apply the new tags to.
It must be an exact version of the dataset to succeed. You can
use the read_dataset_version method to find the exact version
to apply the tags to.
Args:
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to update.
dataset_name (Optional[str]):
The name of the dataset to update.
as_of (datetime.datetime):
The timestamp of the dataset to apply the new tags to.
tag (str):
The new tag to apply to the dataset.
Returns:
None
Examples:
.. code-block:: python
dataset_name = "my-dataset"
# Get the version of a dataset <= a given timestamp
dataset_version = client.read_dataset_version(
dataset_name=dataset_name, as_of=datetime.datetime(2024, 1, 1)
)
# Assign that version a new tag
client.update_dataset_tags(
dataset_name="my-dataset",
as_of=dataset_version.as_of,
tag="prod",
)
"""
if dataset_name is not None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
if dataset_id is None:
raise ValueError("Must provide either dataset name or ID")
response = self.request_with_retries(
"PUT",
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/tags",
headers=self._headers,
json={
"as_of": as_of.isoformat(),
"tag": tag,
},
)
ls_utils.raise_for_status_with_text(response)
def list_dataset_versions(
self,
*,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
search: Optional[str] = None,
limit: Optional[int] = None,
) -> Iterator[ls_schemas.DatasetVersion]:
"""List dataset versions.
Args:
dataset_id (Optional[Union[UUID, str]]): The ID of the dataset.
dataset_name (Optional[str]): The name of the dataset.
search (Optional[str]): The search query.
limit (Optional[int]): The maximum number of versions to return.
Yields:
The dataset versions.
"""
if dataset_id is None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
params = {
"search": search,
"limit": min(limit, 100) if limit is not None else 100,
}
for i, version in enumerate(
self._get_paginated_list(
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/versions",
params=params,
)
):
yield ls_schemas.DatasetVersion(**version)
if limit is not None and i + 1 >= limit:
break
def read_dataset_version(
self,
*,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
as_of: Optional[datetime.datetime] = None,
tag: Optional[str] = None,
) -> ls_schemas.DatasetVersion:
"""Get dataset version by as_of or exact tag.
Ues this to resolve the nearest version to a given timestamp or for a given tag.
Args:
dataset_id (Optional[ID_TYPE]): The ID of the dataset.
dataset_name (Optional[str]): The name of the dataset.
as_of (Optional[datetime.datetime]): The timestamp of the dataset
to retrieve.
tag (Optional[str]): The tag of the dataset to retrieve.
Returns:
DatasetVersion: The dataset version.
Examples:
.. code-block:: python
# Get the latest version of a dataset
client.read_dataset_version(dataset_name="my-dataset", tag="latest")
# Get the version of a dataset <= a given timestamp
client.read_dataset_version(
dataset_name="my-dataset",
as_of=datetime.datetime(2024, 1, 1),
)
# Get the version of a dataset with a specific tag
client.read_dataset_version(dataset_name="my-dataset", tag="prod")
"""
if dataset_id is None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
if (as_of and tag) or (as_of is None and tag is None):
raise ValueError("Exactly one of as_of and tag must be specified.")
response = self.request_with_retries(
"GET",
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/version",
params={"as_of": as_of, "tag": tag},
)
return ls_schemas.DatasetVersion(**response.json())
def clone_public_dataset(
self,
token_or_url: str,
*,
source_api_url: Optional[str] = None,
dataset_name: Optional[str] = None,
) -> ls_schemas.Dataset:
"""Clone a public dataset to your own langsmith tenant.
This operation is idempotent. If you already have a dataset with the given name,
this function will do nothing.
Args:
token_or_url (str): The token of the public dataset to clone.
source_api_url (Optional[str]): The URL of the langsmith server where the data is hosted.
Defaults to the API URL of your current client.
dataset_name (Optional[str]): The name of the dataset to create in your tenant.
Defaults to the name of the public dataset.
Returns:
Dataset: The cloned dataset.
"""
source_api_url = source_api_url or self.api_url
source_api_url, token_uuid = _parse_token_or_url(token_or_url, source_api_url)
source_client = Client(
# Placeholder API key not needed anymore in most cases, but
# some private deployments may have API key-based rate limiting
# that would cause this to fail if we provide no value.
api_url=source_api_url,
api_key="placeholder",
)
ds = source_client.read_shared_dataset(token_uuid)
dataset_name = dataset_name or ds.name
try:
ds = self.read_dataset(dataset_name=dataset_name)
logger.info(
f"Dataset {dataset_name} already exists in your tenant. Skipping."
)
return ds
except ls_utils.LangSmithNotFoundError:
pass
try:
# Fetch examples first
examples = list(source_client.list_shared_examples(token_uuid))
dataset = self.create_dataset(
dataset_name=dataset_name,
description=ds.description,
data_type=ds.data_type or ls_schemas.DataType.kv,
inputs_schema=ds.inputs_schema,
outputs_schema=ds.outputs_schema,
transformations=ds.transformations,
)
try:
self.create_examples(
inputs=[e.inputs for e in examples],
outputs=[e.outputs for e in examples],
dataset_id=dataset.id,
)
except BaseException as e:
# Let's not do automatic clean up for now in case there might be
# some other reasons why create_examples fails (i.e., not network issue
# or keyboard interrupt).
# The risk is that this is an existing dataset that has valid examples
# populated from another source so we don't want to delete it.
logger.error(
f"An error occurred while creating dataset {dataset_name}. "
"You should delete it manually."
)
raise e
finally:
del source_client
return dataset
def _get_data_type(self, dataset_id: ID_TYPE) -> ls_schemas.DataType:
dataset = self.read_dataset(dataset_id=dataset_id)
return dataset.data_type
@ls_utils.xor_args(("dataset_id", "dataset_name"))
def create_llm_example(
self,
prompt: str,
generation: Optional[str] = None,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
created_at: Optional[datetime.datetime] = None,
) -> ls_schemas.Example:
"""Add an example (row) to an LLM-type dataset.
Args:
prompt (str):
The input prompt for the example.
generation (Optional[str]):
The output generation for the example.
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset.
dataset_name (Optional[str]):
The name of the dataset.
created_at (Optional[datetime.datetime]):
The creation timestamp of the example.
Returns:
Example: The created example
"""
return self.create_example(
inputs={"input": prompt},
outputs={"output": generation},
dataset_id=dataset_id,
dataset_name=dataset_name,
created_at=created_at,
)
@ls_utils.xor_args(("dataset_id", "dataset_name"))
def create_chat_example(
self,
messages: List[Union[Mapping[str, Any], ls_schemas.BaseMessageLike]],
generations: Optional[
Union[Mapping[str, Any], ls_schemas.BaseMessageLike]
] = None,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
created_at: Optional[datetime.datetime] = None,
) -> ls_schemas.Example:
"""Add an example (row) to a Chat-type dataset.
Args:
messages (List[Union[Mapping[str, Any], BaseMessageLike]]):
The input messages for the example.
generations (Optional[Union[Mapping[str, Any], BaseMessageLike]]):
The output messages for the example.
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset.
dataset_name (Optional[str]):
The name of the dataset.
created_at (Optional[datetime.datetime]):
The creation timestamp of the example.
Returns:
Example: The created example
"""
final_input = []
for message in messages:
if ls_utils.is_base_message_like(message):
final_input.append(
ls_utils.convert_langchain_message(
cast(ls_schemas.BaseMessageLike, message)
)
)
else:
final_input.append(cast(dict, message))
final_generations = None
if generations is not None:
if ls_utils.is_base_message_like(generations):
final_generations = ls_utils.convert_langchain_message(
cast(ls_schemas.BaseMessageLike, generations)
)
else:
final_generations = cast(dict, generations)
return self.create_example(
inputs={"input": final_input},
outputs=(
{"output": final_generations} if final_generations is not None else None
),
dataset_id=dataset_id,
dataset_name=dataset_name,
created_at=created_at,
)
def create_example_from_run(
self,
run: ls_schemas.Run,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
created_at: Optional[datetime.datetime] = None,
) -> ls_schemas.Example:
"""Add an example (row) to a dataset from a run.
Args:
run (Run): The run to create an example from.
dataset_id (Optional[Union[UUID, str]]): The ID of the dataset.
dataset_name (Optional[str]): The name of the dataset.
created_at (Optional[datetime.datetime]): The creation timestamp of the example.
Returns:
Example: The created example
"""
if dataset_id is None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
dataset_name = None # Nested call expects only 1 defined
dataset_type = self._get_data_type_cached(dataset_id)
if dataset_type == ls_schemas.DataType.llm:
if run.run_type != "llm":
raise ValueError(
f"Run type {run.run_type} is not supported"
" for dataset of type 'LLM'"
)
try:
prompt = ls_utils.get_prompt_from_inputs(run.inputs)
except ValueError:
raise ValueError(
"Error converting LLM run inputs to prompt for run"
f" {run.id} with inputs {run.inputs}"
)
inputs: Dict[str, Any] = {"input": prompt}
if not run.outputs:
outputs: Optional[Dict[str, Any]] = None
else:
try:
generation = ls_utils.get_llm_generation_from_outputs(run.outputs)
except ValueError:
raise ValueError(
"Error converting LLM run outputs to generation for run"
f" {run.id} with outputs {run.outputs}"
)
outputs = {"output": generation}
elif dataset_type == ls_schemas.DataType.chat:
if run.run_type != "llm":
raise ValueError(
f"Run type {run.run_type} is not supported"
" for dataset of type 'chat'"
)
try:
inputs = {"input": ls_utils.get_messages_from_inputs(run.inputs)}
except ValueError:
raise ValueError(
"Error converting LLM run inputs to chat messages for run"
f" {run.id} with inputs {run.inputs}"
)
if not run.outputs:
outputs = None
else:
try:
outputs = {
"output": ls_utils.get_message_generation_from_outputs(
run.outputs
)
}
except ValueError:
raise ValueError(
"Error converting LLM run outputs to chat generations"
f" for run {run.id} with outputs {run.outputs}"
)
elif dataset_type == ls_schemas.DataType.kv:
# Anything goes
inputs = run.inputs
outputs = run.outputs
else:
raise ValueError(f"Dataset type {dataset_type} not recognized.")
return self.create_example(
inputs=inputs,
outputs=outputs,
dataset_id=dataset_id,
dataset_name=dataset_name,
created_at=created_at,
)
def _prepare_multipart_data(
self,
examples: Union[
List[ls_schemas.ExampleUploadWithAttachments]
| List[ls_schemas.ExampleUpsertWithAttachments]
| List[ls_schemas.ExampleUpdateWithAttachments],
],
include_dataset_id: bool = False,
) -> Tuple[Any, bytes]:
parts: List[MultipartPart] = []
if include_dataset_id:
if not isinstance(examples[0], ls_schemas.ExampleUpsertWithAttachments):
raise ValueError(
"The examples must be of type ExampleUpsertWithAttachments"
" if include_dataset_id is True"
)
dataset_id = examples[0].dataset_id
for example in examples:
if (
not isinstance(example, ls_schemas.ExampleUploadWithAttachments)
and not isinstance(example, ls_schemas.ExampleUpsertWithAttachments)
and not isinstance(example, ls_schemas.ExampleUpdateWithAttachments)
):
raise ValueError(
"The examples must be of type ExampleUploadWithAttachments"
" or ExampleUpsertWithAttachments"
" or ExampleUpdateWithAttachments"
)
if example.id is not None:
example_id = str(example.id)
else:
example_id = str(uuid.uuid4())
if isinstance(example, ls_schemas.ExampleUpdateWithAttachments):
created_at = None
else:
created_at = example.created_at
example_body = {
**({"dataset_id": dataset_id} if include_dataset_id else {}),
**({"created_at": created_at} if created_at is not None else {}),
}
if example.metadata is not None:
example_body["metadata"] = example.metadata
if example.split is not None:
example_body["split"] = example.split
valb = _dumps_json(example_body)
parts.append(
(
f"{example_id}",
(
None,
valb,
"application/json",
{},
),
)
)
inputsb = _dumps_json(example.inputs)
parts.append(
(
f"{example_id}.inputs",
(
None,
inputsb,
"application/json",
{},
),
)
)
if example.outputs:
outputsb = _dumps_json(example.outputs)
parts.append(
(
f"{example_id}.outputs",
(
None,
outputsb,
"application/json",
{},
),
)
)
if example.attachments:
for name, attachment in example.attachments.items():
if isinstance(attachment, tuple):
if isinstance(attachment[1], Path):
mime_type, file_path = attachment
file_size = os.path.getsize(file_path)
parts.append(
(
f"{example_id}.attachment.{name}",
(
None,
open(file_path, "rb"), # type: ignore[arg-type]
f"{mime_type}; length={file_size}",
{},
),
)
)
else:
mime_type, data = attachment
parts.append(
(
f"{example_id}.attachment.{name}",
(
None,
data,
f"{mime_type}; length={len(data)}",
{},
),
)
)
else:
parts.append(
(
f"{example_id}.attachment.{name}",
(
None,
attachment.data,
f"{attachment.mime_type}; length={len(attachment.data)}",
{},
),
)
)
if (
isinstance(example, ls_schemas.ExampleUpdateWithAttachments)
and example.attachments_operations
):
attachments_operationsb = _dumps_json(example.attachments_operations)
parts.append(
(
f"{example_id}.attachments_operations",
(
None,
attachments_operationsb,
"application/json",
{},
),
)
)
encoder = rqtb_multipart.MultipartEncoder(parts, boundary=_BOUNDARY)
if encoder.len <= 20_000_000: # ~20 MB
data = encoder.to_string()
else:
data = encoder
return encoder, data
def update_examples_multipart(
self,
*,
dataset_id: ID_TYPE,
updates: Optional[List[ls_schemas.ExampleUpdateWithAttachments]] = None,
) -> ls_schemas.UpsertExamplesResponse:
"""Update examples using multipart.
Args:
dataset_id (Union[UUID, str]): The ID of the dataset to update.
updates (Optional[List[ExampleUpdateWithAttachments]]): The updates to apply to the examples.
Raises:
ValueError: If the multipart examples endpoint is not enabled.
"""
if not (self.info.instance_flags or {}).get(
"dataset_examples_multipart_enabled", False
):
raise ValueError(
"Your LangSmith version does not allow using the multipart examples endpoint, please update to the latest version."
)
if updates is None:
updates = []
encoder, data = self._prepare_multipart_data(updates, include_dataset_id=False)
response = self.request_with_retries(
"PATCH",
f"/v1/platform/datasets/{dataset_id}/examples",
request_kwargs={
"data": data,
"headers": {
**self._headers,
"Content-Type": encoder.content_type,
},
},
)
ls_utils.raise_for_status_with_text(response)
return response.json()
def upload_examples_multipart(
self,
*,
dataset_id: ID_TYPE,
uploads: Optional[List[ls_schemas.ExampleUploadWithAttachments]] = None,
) -> ls_schemas.UpsertExamplesResponse:
"""Upload examples using multipart.
Args:
dataset_id (Union[UUID, str]): The ID of the dataset to upload to.
uploads (Optional[List[ExampleUploadWithAttachments]]): The examples to upload.
Returns:
ls_schemas.UpsertExamplesResponse: The count and ids of the successfully uploaded examples
Raises:
ValueError: If the multipart examples endpoint is not enabled.
"""
if not (self.info.instance_flags or {}).get(
"dataset_examples_multipart_enabled", False
):
raise ValueError(
"Your LangSmith version does not allow using the multipart examples endpoint, please update to the latest version."
)
if uploads is None:
uploads = []
encoder, data = self._prepare_multipart_data(uploads, include_dataset_id=False)
response = self.request_with_retries(
"POST",
f"/v1/platform/datasets/{dataset_id}/examples",
request_kwargs={
"data": data,
"headers": {
**self._headers,
"Content-Type": encoder.content_type,
},
},
)
ls_utils.raise_for_status_with_text(response)
return response.json()
def upsert_examples_multipart(
self,
*,
upserts: Optional[List[ls_schemas.ExampleUpsertWithAttachments]] = None,
) -> ls_schemas.UpsertExamplesResponse:
"""Upsert examples.
.. deprecated:: 0.1.0
This method is deprecated. Use :func:`langsmith.upload_examples_multipart` instead.
""" # noqa: E501
if not (self.info.instance_flags or {}).get(
"examples_multipart_enabled", False
):
raise ValueError(
"Your LangSmith version does not allow using the multipart examples endpoint, please update to the latest version."
)
if upserts is None:
upserts = []
encoder, data = self._prepare_multipart_data(upserts, include_dataset_id=True)
response = self.request_with_retries(
"POST",
"/v1/platform/examples/multipart",
request_kwargs={
"data": data,
"headers": {
**self._headers,
"Content-Type": encoder.content_type,
},
},
)
ls_utils.raise_for_status_with_text(response)
return response.json()
def create_examples(
self,
*,
inputs: Sequence[Mapping[str, Any]],
outputs: Optional[Sequence[Optional[Mapping[str, Any]]]] = None,
metadata: Optional[Sequence[Optional[Mapping[str, Any]]]] = None,
splits: Optional[Sequence[Optional[str | List[str]]]] = None,
source_run_ids: Optional[Sequence[Optional[ID_TYPE]]] = None,
ids: Optional[Sequence[Optional[ID_TYPE]]] = None,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
**kwargs: Any,
) -> None:
"""Create examples in a dataset.
Args:
inputs (Sequence[Mapping[str, Any]]):
The input values for the examples.
outputs (Optional[Sequence[Optional[Mapping[str, Any]]]]):
The output values for the examples.
metadata (Optional[Sequence[Optional[Mapping[str, Any]]]]):
The metadata for the examples.
splits (Optional[Sequence[Optional[str | List[str]]]]):
The splits for the examples, which are divisions
of your dataset such as 'train', 'test', or 'validation'.
source_run_ids (Optional[Sequence[Optional[Union[UUID, str]]]]):
The IDs of the source runs associated with the examples.
ids (Optional[Sequence[Union[UUID, str]]]):
The IDs of the examples.
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to create the examples in.
dataset_name (Optional[str]):
The name of the dataset to create the examples in.
**kwargs: Any: Additional keyword arguments are ignored.
Raises:
ValueError: If neither dataset_id nor dataset_name is provided.
Returns:
None
"""
if dataset_id is None and dataset_name is None:
raise ValueError("Either dataset_id or dataset_name must be provided.")
if dataset_id is None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
sequence_args = {
"outputs": outputs,
"metadata": metadata,
"splits": splits,
"ids": ids,
"source_run_ids": source_run_ids,
}
# Since inputs are required, we will check against them
input_len = len(inputs)
for arg_name, arg_value in sequence_args.items():
if arg_value is not None and len(arg_value) != input_len:
raise ValueError(
f"Length of {arg_name} ({len(arg_value)}) does not match"
f" length of inputs ({input_len})"
)
examples = [
{
"inputs": in_,
"outputs": out_,
"dataset_id": dataset_id,
"metadata": metadata_,
"split": split_,
"id": id_ or str(uuid.uuid4()),
"source_run_id": source_run_id_,
}
for in_, out_, metadata_, split_, id_, source_run_id_ in zip(
inputs,
outputs or [None] * len(inputs),
metadata or [None] * len(inputs),
splits or [None] * len(inputs),
ids or [None] * len(inputs),
source_run_ids or [None] * len(inputs),
)
]
response = self.request_with_retries(
"POST",
"/examples/bulk",
headers={**self._headers, "Content-Type": "application/json"},
data=_dumps_json(examples),
)
ls_utils.raise_for_status_with_text(response)
@ls_utils.xor_args(("dataset_id", "dataset_name"))
def create_example(
self,
inputs: Mapping[str, Any],
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
created_at: Optional[datetime.datetime] = None,
outputs: Optional[Mapping[str, Any]] = None,
metadata: Optional[Mapping[str, Any]] = None,
split: Optional[str | List[str]] = None,
example_id: Optional[ID_TYPE] = None,
source_run_id: Optional[ID_TYPE] = None,
) -> ls_schemas.Example:
"""Create a dataset example in the LangSmith API.
Examples are rows in a dataset, containing the inputs
and expected outputs (or other reference information)
for a model or chain.
Args:
inputs (Mapping[str, Any]):
The input values for the example.
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to create the example in.
dataset_name (Optional[str]):
The name of the dataset to create the example in.
created_at (Optional[datetime.datetime]):
The creation timestamp of the example.
outputs (Optional[Mapping[str, Any]]):
The output values for the example.
metadata (Optional[Mapping[str, Any]]):
The metadata for the example.
split (Optional[str | List[str]]):
The splits for the example, which are divisions
of your dataset such as 'train', 'test', or 'validation'.
example_id (Optional[Union[UUID, str]]):
The ID of the example to create. If not provided, a new
example will be created.
source_run_id (Optional[Union[UUID, str]]):
The ID of the source run associated with this example.
Returns:
Example: The created example.
"""
if dataset_id is None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
data = {
"inputs": inputs,
"outputs": outputs,
"dataset_id": dataset_id,
"metadata": metadata,
"split": split,
"source_run_id": source_run_id,
}
if created_at:
data["created_at"] = created_at.isoformat()
data["id"] = example_id or str(uuid.uuid4())
response = self.request_with_retries(
"POST",
"/examples",
headers={**self._headers, "Content-Type": "application/json"},
data=_dumps_json({k: v for k, v in data.items() if v is not None}),
)
ls_utils.raise_for_status_with_text(response)
result = response.json()
return ls_schemas.Example(
**result,
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
def read_example(
self, example_id: ID_TYPE, *, as_of: Optional[datetime.datetime] = None
) -> ls_schemas.Example:
"""Read an example from the LangSmith API.
Args:
example_id (Union[UUID, str]): The ID of the example to read.
as_of (Optional[datetime.datetime]): The dataset version tag OR
timestamp to retrieve the example as of.
Response examples will only be those that were present at the time
of the tagged (or timestamped) version.
Returns:
Example: The example.
"""
response = self.request_with_retries(
"GET",
f"/examples/{_as_uuid(example_id, 'example_id')}",
params={
"as_of": as_of.isoformat() if as_of else None,
},
)
example = response.json()
attachments = {}
if example.get("attachment_urls"):
for key, value in example["attachment_urls"].items():
response = requests.get(value["presigned_url"], stream=True)
response.raise_for_status()
reader = io.BytesIO(response.content)
attachments[key.removeprefix("attachment.")] = {
"presigned_url": value["presigned_url"],
"reader": reader,
"mime_type": value.get("mime_type"),
}
return ls_schemas.Example(
**{k: v for k, v in example.items() if k != "attachment_urls"},
attachments=attachments,
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
def list_examples(
self,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
example_ids: Optional[Sequence[ID_TYPE]] = None,
as_of: Optional[Union[datetime.datetime, str]] = None,
splits: Optional[Sequence[str]] = None,
inline_s3_urls: bool = True,
*,
offset: int = 0,
limit: Optional[int] = None,
metadata: Optional[dict] = None,
filter: Optional[str] = None,
include_attachments: bool = False,
**kwargs: Any,
) -> Iterator[ls_schemas.Example]:
r"""Retrieve the example rows of the specified dataset.
Args:
dataset_id (Optional[Union[UUID, str]]): The ID of the dataset to filter by.
Defaults to None.
dataset_name (Optional[str]): The name of the dataset to filter by.
Defaults to None.
example_ids (Optional[Sequence[Union[UUID, str]]): The IDs of the examples to filter by.
Defaults to None.
as_of (Optional[Union[datetime.datetime, str]]): The dataset version tag OR
timestamp to retrieve the examples as of.
Response examples will only be those that were present at the time
of the tagged (or timestamped) version.
splits (Optional[Sequence[str]]): A list of dataset splits, which are
divisions of your dataset such as 'train', 'test', or 'validation'.
Returns examples only from the specified splits.
inline_s3_urls (bool, default=True): Whether to inline S3 URLs.
Defaults to True.
offset (int, default=0): The offset to start from. Defaults to 0.
limit (Optional[int]): The maximum number of examples to return.
metadata (Optional[dict]): A dictionary of metadata to filter by.
filter (Optional[str]): A structured fileter string to apply to
the examples.
include_attachments (bool, default=False): Whether to include the
attachments in the response. Defaults to False.
**kwargs (Any): Additional keyword arguments are ignored.
Yields:
The examples.
Examples:
List all examples for a dataset:
.. code-block:: python
from langsmith import Client
client = Client()
# By Dataset ID
examples = client.list_examples(
dataset_id="c9ace0d8-a82c-4b6c-13d2-83401d68e9ab"
)
# By Dataset Name
examples = client.list_examples(dataset_name="My Test Dataset")
List examples by id
.. code-block:: python
example_ids = [
"734fc6a0-c187-4266-9721-90b7a025751a",
"d6b4c1b9-6160-4d63-9b61-b034c585074f",
"4d31df4e-f9c3-4a6e-8b6c-65701c2fed13",
]
examples = client.list_examples(example_ids=example_ids)
List examples by metadata
.. code-block:: python
examples = client.list_examples(
dataset_name=dataset_name, metadata={"foo": "bar"}
)
List examples by structured filter
.. code-block:: python
examples = client.list_examples(
dataset_name=dataset_name,
filter='and(not(has(metadata, \'{"foo": "bar"}\')), exists(metadata, "tenant_id"))',
)
"""
params: Dict[str, Any] = {
**kwargs,
"offset": offset,
"id": example_ids,
"as_of": (
as_of.isoformat() if isinstance(as_of, datetime.datetime) else as_of
),
"splits": splits,
"inline_s3_urls": inline_s3_urls,
"limit": min(limit, 100) if limit is not None else 100,
"filter": filter,
}
if metadata is not None:
params["metadata"] = _dumps_json(metadata)
if dataset_id is not None:
params["dataset"] = dataset_id
elif dataset_name is not None:
dataset_id = self.read_dataset(dataset_name=dataset_name).id
params["dataset"] = dataset_id
else:
pass
if include_attachments:
params["select"] = ["attachment_urls", "outputs", "metadata"]
for i, example in enumerate(
self._get_paginated_list("/examples", params=params)
):
attachments = {}
if example.get("attachment_urls"):
for key, value in example["attachment_urls"].items():
response = requests.get(value["presigned_url"], stream=True)
response.raise_for_status()
reader = io.BytesIO(response.content)
attachments[key.removeprefix("attachment.")] = {
"presigned_url": value["presigned_url"],
"reader": reader,
"mime_type": value.get("mime_type"),
}
yield ls_schemas.Example(
**{k: v for k, v in example.items() if k != "attachment_urls"},
attachments=attachments,
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
if limit is not None and i + 1 >= limit:
break
@warn_beta
def index_dataset(
self,
*,
dataset_id: ID_TYPE,
tag: str = "latest",
**kwargs: Any,
) -> None:
"""Enable dataset indexing. Examples are indexed by their inputs.
This enables searching for similar examples by inputs with
``client.similar_examples()``.
Args:
dataset_id (Union[UUID, str]): The ID of the dataset to index.
tag (Optional[str]): The version of the dataset to index. If 'latest'
then any updates to the dataset (additions, updates, deletions of
examples) will be reflected in the index.
**kwargs (Any): Additional keyword arguments to pass as part of request body.
Returns:
None
""" # noqa: E501
dataset_id = _as_uuid(dataset_id, "dataset_id")
resp = self.request_with_retries(
"POST",
f"/datasets/{dataset_id}/index",
headers=self._headers,
data=json.dumps({"tag": tag, **kwargs}),
)
ls_utils.raise_for_status_with_text(resp)
# NOTE: dataset_name arg explicitly not supported to avoid extra API calls.
@warn_beta
def similar_examples(
self,
inputs: dict,
/,
*,
limit: int,
dataset_id: ID_TYPE,
filter: Optional[str] = None,
**kwargs: Any,
) -> List[ls_schemas.ExampleSearch]:
r"""Retrieve the dataset examples whose inputs best match the current inputs.
**Note**: Must have few-shot indexing enabled for the dataset. See
`client.index_dataset()`.
Args:
inputs (dict): The inputs to use as a search query. Must match the dataset
input schema. Must be JSON serializable.
limit (int): The maximum number of examples to return.
dataset_id (Union[UUID, str]): The ID of the dataset to search over.
filter (Optional[str]): A filter string to apply to the search results. Uses
the same syntax as the `filter` parameter in `list_runs()`. Only a subset
of operations are supported. Defaults to None.
For example, you can use ``and(eq(metadata.some_tag, 'some_value'), neq(metadata.env, 'dev'))``
to filter only examples where some_tag has some_value, and the environment is not dev.
**kwargs: Additional keyword arguments to pass as part of request body.
Returns:
list[ExampleSearch]: List of ExampleSearch objects.
Examples:
.. code-block:: python
from langsmith import Client
client = Client()
client.similar_examples(
{"question": "When would i use the runnable generator"},
limit=3,
dataset_id="...",
)
.. code-block:: python
[
ExampleSearch(
inputs={
"question": "How do I cache a Chat model? What caches can I use?"
},
outputs={
"answer": "You can use LangChain's caching layer for Chat Models. This can save you money by reducing the number of API calls you make to the LLM provider, if you're often requesting the same completion multiple times, and speed up your application.\n\nfrom langchain.cache import InMemoryCache\nlangchain.llm_cache = InMemoryCache()\n\n# The first time, it is not yet in cache, so it should take longer\nllm.predict('Tell me a joke')\n\nYou can also use SQLite Cache which uses a SQLite database:\n\nrm .langchain.db\n\nfrom langchain.cache import SQLiteCache\nlangchain.llm_cache = SQLiteCache(database_path=\".langchain.db\")\n\n# The first time, it is not yet in cache, so it should take longer\nllm.predict('Tell me a joke') \n"
},
metadata=None,
id=UUID("b2ddd1c4-dff6-49ae-8544-f48e39053398"),
dataset_id=UUID("01b6ce0f-bfb6-4f48-bbb8-f19272135d40"),
),
ExampleSearch(
inputs={"question": "What's a runnable lambda?"},
outputs={
"answer": "A runnable lambda is an object that implements LangChain's `Runnable` interface and runs a callbale (i.e., a function). Note the function must accept a single argument."
},
metadata=None,
id=UUID("f94104a7-2434-4ba7-8293-6a283f4860b4"),
dataset_id=UUID("01b6ce0f-bfb6-4f48-bbb8-f19272135d40"),
),
ExampleSearch(
inputs={"question": "Show me how to use RecursiveURLLoader"},
outputs={
"answer": 'The RecursiveURLLoader comes from the langchain.document_loaders.recursive_url_loader module. Here\'s an example of how to use it:\n\nfrom langchain.document_loaders.recursive_url_loader import RecursiveUrlLoader\n\n# Create an instance of RecursiveUrlLoader with the URL you want to load\nloader = RecursiveUrlLoader(url="https://example.com")\n\n# Load all child links from the URL page\nchild_links = loader.load()\n\n# Print the child links\nfor link in child_links:\n print(link)\n\nMake sure to replace "https://example.com" with the actual URL you want to load. The load() method returns a list of child links found on the URL page. You can iterate over this list to access each child link.'
},
metadata=None,
id=UUID("0308ea70-a803-4181-a37d-39e95f138f8c"),
dataset_id=UUID("01b6ce0f-bfb6-4f48-bbb8-f19272135d40"),
),
]
"""
dataset_id = _as_uuid(dataset_id, "dataset_id")
req = {
"inputs": inputs,
"limit": limit,
**kwargs,
}
if filter is not None:
req["filter"] = filter
resp = self.request_with_retries(
"POST",
f"/datasets/{dataset_id}/search",
headers=self._headers,
data=json.dumps(req),
)
ls_utils.raise_for_status_with_text(resp)
examples = []
for ex in resp.json()["examples"]:
examples.append(ls_schemas.ExampleSearch(**ex, dataset_id=dataset_id))
return examples
def update_example(
self,
example_id: ID_TYPE,
*,
inputs: Optional[Dict[str, Any]] = None,
outputs: Optional[Mapping[str, Any]] = None,
metadata: Optional[Dict] = None,
split: Optional[str | List[str]] = None,
dataset_id: Optional[ID_TYPE] = None,
attachments_operations: Optional[ls_schemas.AttachmentsOperations] = None,
) -> Dict[str, Any]:
"""Update a specific example.
Args:
example_id (Union[UUID, str]):
The ID of the example to update.
inputs (Optional[Dict[str, Any]]):
The input values to update.
outputs (Optional[Mapping[str, Any]]):
The output values to update.
metadata (Optional[Dict]):
The metadata to update.
split (Optional[str | List[str]]):
The dataset split to update, such as
'train', 'test', or 'validation'.
dataset_id (Optional[Union[UUID, str]]):
The ID of the dataset to update.
attachments_operations (Optional[AttachmentsOperations]):
The attachments operations to perform.
Returns:
Dict[str, Any]: The updated example.
"""
if attachments_operations is not None:
if not (self.info.instance_flags or {}).get(
"dataset_examples_multipart_enabled", False
):
raise ValueError(
"Your LangSmith version does not allow using the attachment operations, please update to the latest version."
)
example = dict(
inputs=inputs,
outputs=outputs,
dataset_id=dataset_id,
metadata=metadata,
split=split,
attachments_operations=attachments_operations,
)
response = self.request_with_retries(
"PATCH",
f"/examples/{_as_uuid(example_id, 'example_id')}",
headers={**self._headers, "Content-Type": "application/json"},
data=_dumps_json({k: v for k, v in example.items() if v is not None}),
)
ls_utils.raise_for_status_with_text(response)
return response.json()
def update_examples(
self,
*,
example_ids: Sequence[ID_TYPE],
inputs: Optional[Sequence[Optional[Dict[str, Any]]]] = None,
outputs: Optional[Sequence[Optional[Mapping[str, Any]]]] = None,
metadata: Optional[Sequence[Optional[Dict]]] = None,
splits: Optional[Sequence[Optional[str | List[str]]]] = None,
dataset_ids: Optional[Sequence[Optional[ID_TYPE]]] = None,
attachments_operations: Optional[
Sequence[Optional[ls_schemas.AttachmentsOperations]]
] = None,
) -> Dict[str, Any]:
"""Update multiple examples.
Args:
example_ids (Sequence[Union[UUID, str]]):
The IDs of the examples to update.
inputs (Optional[Sequence[Optional[Dict[str, Any]]]):
The input values for the examples.
outputs (Optional[Sequence[Optional[Mapping[str, Any]]]]):
The output values for the examples.
metadata (Optional[Sequence[Optional[Mapping[str, Any]]]]):
The metadata for the examples.
splits (Optional[Sequence[Optional[str | List[str]]]]):
The splits for the examples, which are divisions
of your dataset such as 'train', 'test', or 'validation'.
dataset_ids (Optional[Sequence[Optional[Union[UUID, str]]]]):
The IDs of the datasets to move the examples to.
attachments_operations (Optional[Sequence[Optional[ls_schemas.AttachmentsOperations]]):
The operations to perform on the attachments.
Returns:
Dict[str, Any]: The response from the server (specifies the number of examples updated).
"""
if attachments_operations is not None:
if not (self.info.instance_flags or {}).get(
"dataset_examples_multipart_enabled", False
):
raise ValueError(
"Your LangSmith version does not allow using the attachment operations, please update to the latest version."
)
sequence_args = {
"inputs": inputs,
"outputs": outputs,
"metadata": metadata,
"splits": splits,
"dataset_ids": dataset_ids,
"attachments_operations": attachments_operations,
}
# Since inputs are required, we will check against them
examples_len = len(example_ids)
for arg_name, arg_value in sequence_args.items():
if arg_value is not None and len(arg_value) != examples_len:
raise ValueError(
f"Length of {arg_name} ({len(arg_value)}) does not match"
f" length of examples ({examples_len})"
)
examples = [
{
"id": id_,
"inputs": in_,
"outputs": out_,
"dataset_id": dataset_id_,
"metadata": metadata_,
"split": split_,
"attachments_operations": attachments_operations_,
}
for id_, in_, out_, metadata_, split_, dataset_id_, attachments_operations_ in zip(
example_ids,
inputs or [None] * len(example_ids),
outputs or [None] * len(example_ids),
metadata or [None] * len(example_ids),
splits or [None] * len(example_ids),
dataset_ids or [None] * len(example_ids),
attachments_operations or [None] * len(example_ids),
)
]
response = self.request_with_retries(
"PATCH",
"/examples/bulk",
headers={**self._headers, "Content-Type": "application/json"},
data=(
_dumps_json(
[
{k: v for k, v in example.items() if v is not None}
for example in examples
]
)
),
)
ls_utils.raise_for_status_with_text(response)
return response.json()
def delete_example(self, example_id: ID_TYPE) -> None:
"""Delete an example by ID.
Args:
example_id (Union[UUID, str]):
The ID of the example to delete.
Returns:
None
"""
response = self.request_with_retries(
"DELETE",
f"/examples/{_as_uuid(example_id, 'example_id')}",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
def delete_examples(self, example_ids: Sequence[ID_TYPE]) -> None:
"""Delete multiple examples by ID.
Parameters
----------
example_ids : Sequence[ID_TYPE]
The IDs of the examples to delete.
"""
response = self.request_with_retries(
"DELETE",
"/examples",
headers={**self._headers, "Content-Type": "application/json"},
data=_dumps_json(
{
"ids": [
str(_as_uuid(id_, f"example_ids[{i}]"))
for i, id_ in enumerate(example_ids)
]
}
),
)
ls_utils.raise_for_status_with_text(response)
def list_dataset_splits(
self,
*,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
as_of: Optional[Union[str, datetime.datetime]] = None,
) -> List[str]:
"""Get the splits for a dataset.
Args:
dataset_id (Optional[Union[UUID, str]]): The ID of the dataset.
dataset_name (Optional[str]): The name of the dataset.
as_of (Optional[Union[str, datetime.datetime]]): The version
of the dataset to retrieve splits for. Can be a timestamp or a
string tag. Defaults to "latest".
Returns:
List[str]: The names of this dataset's splits.
"""
if dataset_id is None:
if dataset_name is None:
raise ValueError("Must provide dataset name or ID")
dataset_id = self.read_dataset(dataset_name=dataset_name).id
params = {}
if as_of is not None:
params["as_of"] = (
as_of.isoformat() if isinstance(as_of, datetime.datetime) else as_of
)
response = self.request_with_retries(
"GET",
f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/splits",
params=params,
)
ls_utils.raise_for_status_with_text(response)
return response.json()
def update_dataset_splits(
self,
*,
dataset_id: Optional[ID_TYPE] = None,
dataset_name: Optional[str] = None,
split_name: str,
example_ids: List[ID_TYPE],
remove: bool = False,
) -> None:
"""Update the splits for a dataset.
Args:
dataset_id (Optional[Union[UUID, str]]): The ID of the dataset to update.
dataset_name (Optional[str]): The name of the dataset to update.
split_name (str): The name of the split to update.
example_ids (List[Union[UUID, str]]): The IDs of the examples to add to or
remove from the split.
remove (Optional[bool]): If True, remove the examples from the split.
If False, add the examples to the split. Defaults to False.
Returns:
None
"""
if dataset_id is None:
if dataset_name is None:
raise ValueError("Must provide dataset name or ID")
dataset_id = self.read_dataset(dataset_name=dataset_name).id
data = {
"split_name": split_name,
"examples": [
str(_as_uuid(id_, f"example_ids[{i}]"))
for i, id_ in enumerate(example_ids)
],
"remove": remove,
}
response = self.request_with_retries(
"PUT", f"/datasets/{_as_uuid(dataset_id, 'dataset_id')}/splits", json=data
)
ls_utils.raise_for_status_with_text(response)
def _resolve_run_id(
self,
run: Union[ls_schemas.Run, ls_schemas.RunBase, str, uuid.UUID],
load_child_runs: bool,
) -> ls_schemas.Run:
"""Resolve the run ID.
Args:
run (Union[Run, RunBase, str, UUID]):
The run to resolve.
load_child_runs (bool):
Whether to load child runs.
Returns:
Run: The resolved run.
Raises:
TypeError: If the run type is invalid.
"""
if isinstance(run, (str, uuid.UUID)):
run_ = self.read_run(run, load_child_runs=load_child_runs)
else:
run_ = cast(ls_schemas.Run, run)
return run_
def _resolve_example_id(
self,
example: Union[ls_schemas.Example, str, uuid.UUID, dict, None],
run: ls_schemas.Run,
) -> Optional[ls_schemas.Example]:
"""Resolve the example ID.
Args:
example (Optional[Union[Example, str, UUID, dict]]):
The example to resolve.
run (Run):
The run associated with the example.
Returns:
Optional[Example]: The resolved example.
"""
if isinstance(example, (str, uuid.UUID)):
reference_example_ = self.read_example(example)
elif isinstance(example, ls_schemas.Example):
reference_example_ = example
elif isinstance(example, dict):
reference_example_ = ls_schemas.Example(
**example,
_host_url=self._host_url,
_tenant_id=self._get_optional_tenant_id(),
)
elif run.reference_example_id is not None:
reference_example_ = self.read_example(run.reference_example_id)
else:
reference_example_ = None
return reference_example_
def _select_eval_results(
self,
results: Union[
ls_evaluator.EvaluationResult, ls_evaluator.EvaluationResults, dict
],
*,
fn_name: Optional[str] = None,
) -> List[ls_evaluator.EvaluationResult]:
from langsmith.evaluation import evaluator as ls_evaluator # noqa: F811
def _cast_result(
single_result: Union[ls_evaluator.EvaluationResult, dict],
) -> ls_evaluator.EvaluationResult:
if isinstance(single_result, dict):
return ls_evaluator.EvaluationResult(
**{
"key": fn_name,
"comment": single_result.get("reasoning"),
**single_result,
}
)
return single_result
def _is_eval_results(results: Any) -> TypeGuard[ls_evaluator.EvaluationResults]:
return isinstance(results, dict) and "results" in results
if isinstance(results, ls_evaluator.EvaluationResult):
results_ = [results]
elif _is_eval_results(results):
results_ = [_cast_result(r) for r in results["results"]]
elif isinstance(results, dict):
results_ = [_cast_result(cast(dict, results))]
else:
raise ValueError(
f"Invalid evaluation results type: {type(results)}."
" Must be EvaluationResult, EvaluationResults."
)
return results_
def evaluate_run(
self,
run: Union[ls_schemas.Run, ls_schemas.RunBase, str, uuid.UUID],
evaluator: ls_evaluator.RunEvaluator,
*,
source_info: Optional[Dict[str, Any]] = None,
reference_example: Optional[
Union[ls_schemas.Example, str, dict, uuid.UUID]
] = None,
load_child_runs: bool = False,
) -> ls_evaluator.EvaluationResult:
"""Evaluate a run.
Args:
run (Union[Run, RunBase, str, UUID]):
The run to evaluate.
evaluator (RunEvaluator):
The evaluator to use.
source_info (Optional[Dict[str, Any]]):
Additional information about the source of the evaluation to log
as feedback metadata.
reference_example (Optional[Union[Example, str, dict, UUID]]):
The example to use as a reference for the evaluation.
If not provided, the run's reference example will be used.
load_child_runs (bool, default=False):
Whether to load child runs when resolving the run ID.
Returns:
Feedback: The feedback object created by the evaluation.
"""
run_ = self._resolve_run_id(run, load_child_runs=load_child_runs)
reference_example_ = self._resolve_example_id(reference_example, run_)
evaluator_response = evaluator.evaluate_run(
run_,
example=reference_example_,
)
results = self._log_evaluation_feedback(
evaluator_response,
run_,
source_info=source_info,
)
# TODO: Return all results
return results[0]
def _log_evaluation_feedback(
self,
evaluator_response: Union[
ls_evaluator.EvaluationResult, ls_evaluator.EvaluationResults, dict
],
run: Optional[ls_schemas.Run] = None,
source_info: Optional[Dict[str, Any]] = None,
project_id: Optional[ID_TYPE] = None,
*,
_executor: Optional[cf.ThreadPoolExecutor] = None,
) -> List[ls_evaluator.EvaluationResult]:
results = self._select_eval_results(evaluator_response)
def _submit_feedback(**kwargs):
if _executor:
_executor.submit(self.create_feedback, **kwargs)
else:
self.create_feedback(**kwargs)
for res in results:
source_info_ = source_info or {}
if res.evaluator_info:
source_info_ = {**res.evaluator_info, **source_info_}
run_id_ = None
if res.target_run_id:
run_id_ = res.target_run_id
elif run is not None:
run_id_ = run.id
_submit_feedback(
run_id=run_id_,
key=res.key,
score=res.score,
value=res.value,
comment=res.comment,
correction=res.correction,
source_info=source_info_,
source_run_id=res.source_run_id,
feedback_config=cast(
Optional[ls_schemas.FeedbackConfig], res.feedback_config
),
feedback_source_type=ls_schemas.FeedbackSourceType.MODEL,
project_id=project_id,
extra=res.extra,
trace_id=run.trace_id if run else None,
)
return results
async def aevaluate_run(
self,
run: Union[ls_schemas.Run, str, uuid.UUID],
evaluator: ls_evaluator.RunEvaluator,
*,
source_info: Optional[Dict[str, Any]] = None,
reference_example: Optional[
Union[ls_schemas.Example, str, dict, uuid.UUID]
] = None,
load_child_runs: bool = False,
) -> ls_evaluator.EvaluationResult:
"""Evaluate a run asynchronously.
Args:
run (Union[Run, str, UUID]):
The run to evaluate.
evaluator (RunEvaluator):
The evaluator to use.
source_info (Optional[Dict[str, Any]]):
Additional information about the source of the evaluation to log
as feedback metadata.
reference_example (Optional[Union[Example, str, dict, UUID]]):
The example to use as a reference for the evaluation.
If not provided, the run's reference example will be used.
load_child_runs (bool, default=False)
Whether to load child runs when resolving the run ID.
Returns:
EvaluationResult: The evaluation result object created by the evaluation.
"""
run_ = self._resolve_run_id(run, load_child_runs=load_child_runs)
reference_example_ = self._resolve_example_id(reference_example, run_)
evaluator_response = await evaluator.aevaluate_run(
run_,
example=reference_example_,
)
# TODO: Return all results and use async API
results = self._log_evaluation_feedback(
evaluator_response,
run_,
source_info=source_info,
)
return results[0]
def create_feedback(
self,
run_id: Optional[ID_TYPE],
key: str,
*,
score: Union[float, int, bool, None] = None,
value: Union[str, dict, None] = None,
correction: Union[dict, None] = None,
comment: Union[str, None] = None,
source_info: Optional[Dict[str, Any]] = None,
feedback_source_type: Union[
ls_schemas.FeedbackSourceType, str
] = ls_schemas.FeedbackSourceType.API,
source_run_id: Optional[ID_TYPE] = None,
feedback_id: Optional[ID_TYPE] = None,
feedback_config: Optional[ls_schemas.FeedbackConfig] = None,
stop_after_attempt: int = 10,
project_id: Optional[ID_TYPE] = None,
comparative_experiment_id: Optional[ID_TYPE] = None,
feedback_group_id: Optional[ID_TYPE] = None,
extra: Optional[Dict] = None,
trace_id: Optional[ID_TYPE] = None,
**kwargs: Any,
) -> ls_schemas.Feedback:
"""Create a feedback in the LangSmith API.
Args:
run_id (Optional[Union[UUID, str]]):
The ID of the run to provide feedback for. Either the run_id OR
the project_id must be provided.
key (str):
The name of the metric or 'aspect' this feedback is about.
score (Optional[Union[float, int, bool]]):
The score to rate this run on the metric or aspect.
value (Optional[Union[float, int, bool, str, dict]]):
The display value or non-numeric value for this feedback.
correction (Optional[dict]):
The proper ground truth for this run.
comment (Optional[str]):
A comment about this feedback, such as a justification for the score or
chain-of-thought trajectory for an LLM judge.
source_info (Optional[Dict[str, Any]]):
Information about the source of this feedback.
feedback_source_type (Union[FeedbackSourceType, str]):
The type of feedback source, such as model (for model-generated feedback)
or API.
source_run_id (Optional[Union[UUID, str]]):
The ID of the run that generated this feedback, if a "model" type.
feedback_id (Optional[Union[UUID, str]]):
The ID of the feedback to create. If not provided, a random UUID will be
generated.
feedback_config (Optional[FeedbackConfig]):
The configuration specifying how to interpret feedback with this key.
Examples include continuous (with min/max bounds), categorical,
or freeform.
stop_after_attempt (int, default=10):
The number of times to retry the request before giving up.
project_id (Optional[Union[UUID, str]]):
The ID of the project_id to provide feedback on. One - and only one - of
this and run_id must be provided.
comparative_experiment_id (Optional[Union[UUID, str]]):
If this feedback was logged as a part of a comparative experiment, this
associates the feedback with that experiment.
feedback_group_id (Optional[Union[UUID, str]]):
When logging preferences, ranking runs, or other comparative feedback,
this is used to group feedback together.
extra (Optional[Dict]):
Metadata for the feedback.
trace_id (Optional[Union[UUID, str]]):
The trace ID of the run to provide feedback for. Enables batch ingestion.
**kwargs (Any):
Additional keyword arguments.
Returns:
Feedback: The created feedback object.
"""
if run_id is None and project_id is None:
raise ValueError("One of run_id and project_id must be provided")
if run_id is not None and project_id is not None:
raise ValueError("Only one of run_id and project_id must be provided")
if kwargs:
warnings.warn(
"The following arguments are no longer used in the create_feedback"
f" endpoint: {sorted(kwargs)}",
DeprecationWarning,
)
try:
if not isinstance(feedback_source_type, ls_schemas.FeedbackSourceType):
feedback_source_type = ls_schemas.FeedbackSourceType(
feedback_source_type
)
if feedback_source_type == ls_schemas.FeedbackSourceType.API:
feedback_source: ls_schemas.FeedbackSourceBase = (
ls_schemas.APIFeedbackSource(metadata=source_info)
)
elif feedback_source_type == ls_schemas.FeedbackSourceType.MODEL:
feedback_source = ls_schemas.ModelFeedbackSource(metadata=source_info)
else:
raise ValueError(f"Unknown feedback source type {feedback_source_type}")
feedback_source.metadata = (
feedback_source.metadata if feedback_source.metadata is not None else {}
)
if source_run_id is not None and "__run" not in feedback_source.metadata:
feedback_source.metadata["__run"] = {"run_id": str(source_run_id)}
if feedback_source.metadata and "__run" in feedback_source.metadata:
# Validate that the linked run ID is a valid UUID
# Run info may be a base model or dict.
_run_meta: Union[dict, Any] = feedback_source.metadata["__run"]
if hasattr(_run_meta, "dict") and callable(_run_meta):
_run_meta = _run_meta.dict()
if "run_id" in _run_meta:
_run_meta["run_id"] = str(
_as_uuid(
feedback_source.metadata["__run"]["run_id"],
"feedback_source.metadata['__run']['run_id']",
)
)
feedback_source.metadata["__run"] = _run_meta
feedback = ls_schemas.FeedbackCreate(
id=_ensure_uuid(feedback_id),
# If run_id is None, this is interpreted as session-level
# feedback.
run_id=_ensure_uuid(run_id, accept_null=True),
trace_id=_ensure_uuid(trace_id, accept_null=True),
key=key,
score=score,
value=value,
correction=correction,
comment=comment,
feedback_source=feedback_source,
created_at=datetime.datetime.now(datetime.timezone.utc),
modified_at=datetime.datetime.now(datetime.timezone.utc),
feedback_config=feedback_config,
session_id=_ensure_uuid(project_id, accept_null=True),
comparative_experiment_id=_ensure_uuid(
comparative_experiment_id, accept_null=True
),
feedback_group_id=_ensure_uuid(feedback_group_id, accept_null=True),
extra=extra,
)
use_multipart = (self.info.batch_ingest_config or {}).get(
"use_multipart_endpoint", False
)
if (
use_multipart
and self.info.version # TODO: Remove version check once versions have updated
and ls_utils.is_version_greater_or_equal(self.info.version, "0.8.10")
and self.tracing_queue is not None
and feedback.trace_id is not None
):
serialized_op = serialize_feedback_dict(feedback)
self.tracing_queue.put(
TracingQueueItem(str(feedback.id), serialized_op)
)
else:
feedback_block = _dumps_json(feedback.dict(exclude_none=True))
self.request_with_retries(
"POST",
"/feedback",
request_kwargs={
"data": feedback_block,
},
stop_after_attempt=stop_after_attempt,
retry_on=(ls_utils.LangSmithNotFoundError,),
)
return ls_schemas.Feedback(**feedback.dict())
except Exception as e:
logger.error("Error creating feedback", exc_info=True)
raise e
def update_feedback(
self,
feedback_id: ID_TYPE,
*,
score: Union[float, int, bool, None] = None,
value: Union[float, int, bool, str, dict, None] = None,
correction: Union[dict, None] = None,
comment: Union[str, None] = None,
) -> None:
"""Update a feedback in the LangSmith API.
Args:
feedback_id (Union[UUID, str]):
The ID of the feedback to update.
score (Optional[Union[float, int, bool]]):
The score to update the feedback with.
value (Optional[Union[float, int, bool, str, dict]]):
The value to update the feedback with.
correction (Optional[dict]):
The correction to update the feedback with.
comment (Optional[str]):
The comment to update the feedback with.
Returns:
None
"""
feedback_update: Dict[str, Any] = {}
if score is not None:
feedback_update["score"] = score
if value is not None:
feedback_update["value"] = value
if correction is not None:
feedback_update["correction"] = correction
if comment is not None:
feedback_update["comment"] = comment
response = self.request_with_retries(
"PATCH",
f"/feedback/{_as_uuid(feedback_id, 'feedback_id')}",
headers={**self._headers, "Content-Type": "application/json"},
data=_dumps_json(feedback_update),
)
ls_utils.raise_for_status_with_text(response)
def read_feedback(self, feedback_id: ID_TYPE) -> ls_schemas.Feedback:
"""Read a feedback from the LangSmith API.
Args:
feedback_id (Union[UUID, str]):
The ID of the feedback to read.
Returns:
Feedback: The feedback.
"""
response = self.request_with_retries(
"GET",
f"/feedback/{_as_uuid(feedback_id, 'feedback_id')}",
)
return ls_schemas.Feedback(**response.json())
def list_feedback(
self,
*,
run_ids: Optional[Sequence[ID_TYPE]] = None,
feedback_key: Optional[Sequence[str]] = None,
feedback_source_type: Optional[Sequence[ls_schemas.FeedbackSourceType]] = None,
limit: Optional[int] = None,
**kwargs: Any,
) -> Iterator[ls_schemas.Feedback]:
"""List the feedback objects on the LangSmith API.
Args:
run_ids (Optional[Sequence[Union[UUID, str]]]):
The IDs of the runs to filter by.
feedback_key (Optional[Sequence[str]]):
The feedback key(s) to filter by. Examples: 'correctness'
The query performs a union of all feedback keys.
feedback_source_type (Optional[Sequence[FeedbackSourceType]]):
The type of feedback source, such as model or API.
limit (Optional[int]):
The maximum number of feedback to return.
**kwargs (Any):
Additional keyword arguments.
Yields:
The feedback objects.
"""
params: dict = {
"run": run_ids,
"limit": min(limit, 100) if limit is not None else 100,
**kwargs,
}
if feedback_key is not None:
params["key"] = feedback_key
if feedback_source_type is not None:
params["source"] = feedback_source_type
for i, feedback in enumerate(
self._get_paginated_list("/feedback", params=params)
):
yield ls_schemas.Feedback(**feedback)
if limit is not None and i + 1 >= limit:
break
def delete_feedback(self, feedback_id: ID_TYPE) -> None:
"""Delete a feedback by ID.
Args:
feedback_id (Union[UUID, str]):
The ID of the feedback to delete.
Returns:
None
"""
response = self.request_with_retries(
"DELETE",
f"/feedback/{_as_uuid(feedback_id, 'feedback_id')}",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
def create_feedback_from_token(
self,
token_or_url: Union[str, uuid.UUID],
score: Union[float, int, bool, None] = None,
*,
value: Union[float, int, bool, str, dict, None] = None,
correction: Union[dict, None] = None,
comment: Union[str, None] = None,
metadata: Optional[dict] = None,
) -> None:
"""Create feedback from a presigned token or URL.
Args:
token_or_url (Union[str, uuid.UUID]): The token or URL from which to create
feedback.
score (Optional[Union[float, int, bool]]): The score of the feedback.
Defaults to None.
value (Optional[Union[float, int, bool, str, dict]]): The value of the
feedback. Defaults to None.
correction (Optional[dict]): The correction of the feedback.
Defaults to None.
comment (Optional[str]): The comment of the feedback. Defaults
to None.
metadata (Optional[dict]): Additional metadata for the feedback.
Defaults to None.
Raises:
ValueError: If the source API URL is invalid.
Returns:
None
"""
source_api_url, token_uuid = _parse_token_or_url(
token_or_url, self.api_url, num_parts=1
)
if source_api_url != self.api_url:
raise ValueError(f"Invalid source API URL. {source_api_url}")
response = self.request_with_retries(
"POST",
f"/feedback/tokens/{_as_uuid(token_uuid)}",
data=_dumps_json(
{
"score": score,
"value": value,
"correction": correction,
"comment": comment,
"metadata": metadata,
# TODO: Add ID once the API supports it.
}
),
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
def create_presigned_feedback_token(
self,
run_id: ID_TYPE,
feedback_key: str,
*,
expiration: Optional[datetime.datetime | datetime.timedelta] = None,
feedback_config: Optional[ls_schemas.FeedbackConfig] = None,
feedback_id: Optional[ID_TYPE] = None,
) -> ls_schemas.FeedbackIngestToken:
"""Create a pre-signed URL to send feedback data to.
This is useful for giving browser-based clients a way to upload
feedback data directly to LangSmith without accessing the
API key.
Args:
run_id (Union[UUID, str]):
The ID of the run.
feedback_key (str):
The key of the feedback to create.
expiration (Optional[datetime.datetime | datetime.timedelta]): The expiration time of the pre-signed URL.
Either a datetime or a timedelta offset from now.
Default to 3 hours.
feedback_config (Optional[FeedbackConfig]):
If creating a feedback_key for the first time,
this defines how the metric should be interpreted,
such as a continuous score (w/ optional bounds),
or distribution over categorical values.
feedback_id (Optional[Union[UUID, str]): The ID of the feedback to create. If not provided, a new
feedback will be created.
Returns:
FeedbackIngestToken: The pre-signed URL for uploading feedback data.
"""
body: Dict[str, Any] = {
"run_id": run_id,
"feedback_key": feedback_key,
"feedback_config": feedback_config,
"id": feedback_id or str(uuid.uuid4()),
}
if expiration is None:
body["expires_in"] = ls_schemas.TimeDeltaInput(
days=0,
hours=3,
minutes=0,
)
elif isinstance(expiration, datetime.datetime):
body["expires_at"] = expiration.isoformat()
elif isinstance(expiration, datetime.timedelta):
body["expires_in"] = ls_schemas.TimeDeltaInput(
days=expiration.days,
hours=expiration.seconds // 3600,
minutes=(expiration.seconds // 60) % 60,
)
else:
raise ValueError(f"Unknown expiration type: {type(expiration)}")
response = self.request_with_retries(
"POST",
"/feedback/tokens",
data=_dumps_json(body),
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.FeedbackIngestToken(**response.json())
def create_presigned_feedback_tokens(
self,
run_id: ID_TYPE,
feedback_keys: Sequence[str],
*,
expiration: Optional[datetime.datetime | datetime.timedelta] = None,
feedback_configs: Optional[
Sequence[Optional[ls_schemas.FeedbackConfig]]
] = None,
) -> Sequence[ls_schemas.FeedbackIngestToken]:
"""Create a pre-signed URL to send feedback data to.
This is useful for giving browser-based clients a way to upload
feedback data directly to LangSmith without accessing the
API key.
Args:
run_id (Union[UUID, str]):
The ID of the run.
feedback_keys (Sequence[str]):
The key of the feedback to create.
expiration (Optional[datetime.datetime | datetime.timedelta]): The expiration time of the pre-signed URL.
Either a datetime or a timedelta offset from now.
Default to 3 hours.
feedback_configs (Optional[Sequence[Optional[FeedbackConfig]]]):
If creating a feedback_key for the first time,
this defines how the metric should be interpreted,
such as a continuous score (w/ optional bounds),
or distribution over categorical values.
Returns:
Sequence[FeedbackIngestToken]: The pre-signed URL for uploading feedback data.
"""
# validate
if feedback_configs is not None and len(feedback_keys) != len(feedback_configs):
raise ValueError(
"The length of feedback_keys and feedback_configs must be the same."
)
if not feedback_configs:
feedback_configs = [None] * len(feedback_keys)
# build expiry option
expires_in, expires_at = None, None
if expiration is None:
expires_in = ls_schemas.TimeDeltaInput(
days=0,
hours=3,
minutes=0,
)
elif isinstance(expiration, datetime.datetime):
expires_at = expiration.isoformat()
elif isinstance(expiration, datetime.timedelta):
expires_in = ls_schemas.TimeDeltaInput(
days=expiration.days,
hours=expiration.seconds // 3600,
minutes=(expiration.seconds // 60) % 60,
)
else:
raise ValueError(f"Unknown expiration type: {type(expiration)}")
# assemble body, one entry per key
body = _dumps_json(
[
{
"run_id": run_id,
"feedback_key": feedback_key,
"feedback_config": feedback_config,
"expires_in": expires_in,
"expires_at": expires_at,
}
for feedback_key, feedback_config in zip(
feedback_keys, feedback_configs
)
]
)
def req(api_url: str, api_key: Optional[str]) -> list:
response = self.request_with_retries(
"POST",
f"{api_url}/feedback/tokens",
request_kwargs={
"data": body,
"headers": {
**self._headers,
X_API_KEY: api_key or self.api_key,
},
},
)
ls_utils.raise_for_status_with_text(response)
return response.json()
tokens = []
with cf.ThreadPoolExecutor(max_workers=len(self._write_api_urls)) as executor:
futs = [
executor.submit(req, api_url, api_key)
for api_url, api_key in self._write_api_urls.items()
]
for fut in cf.as_completed(futs):
response = fut.result()
tokens.extend(
[ls_schemas.FeedbackIngestToken(**part) for part in response]
)
return tokens
def list_presigned_feedback_tokens(
self,
run_id: ID_TYPE,
*,
limit: Optional[int] = None,
) -> Iterator[ls_schemas.FeedbackIngestToken]:
"""List the feedback ingest tokens for a run.
Args:
run_id (Union[UUID, str]): The ID of the run to filter by.
limit (Optional[int]): The maximum number of tokens to return.
Yields:
The feedback ingest tokens.
"""
params = {
"run_id": _as_uuid(run_id, "run_id"),
"limit": min(limit, 100) if limit is not None else 100,
}
for i, token in enumerate(
self._get_paginated_list("/feedback/tokens", params=params)
):
yield ls_schemas.FeedbackIngestToken(**token)
if limit is not None and i + 1 >= limit:
break
# Annotation Queue API
def list_annotation_queues(
self,
*,
queue_ids: Optional[List[ID_TYPE]] = None,
name: Optional[str] = None,
name_contains: Optional[str] = None,
limit: Optional[int] = None,
) -> Iterator[ls_schemas.AnnotationQueue]:
"""List the annotation queues on the LangSmith API.
Args:
queue_ids (Optional[List[Union[UUID, str]]]):
The IDs of the queues to filter by.
name (Optional[str]):
The name of the queue to filter by.
name_contains (Optional[str]):
The substring that the queue name should contain.
limit (Optional[int]):
The maximum number of queues to return.
Yields:
The annotation queues.
"""
params: dict = {
"ids": (
[_as_uuid(id_, f"queue_ids[{i}]") for i, id_ in enumerate(queue_ids)]
if queue_ids is not None
else None
),
"name": name,
"name_contains": name_contains,
"limit": min(limit, 100) if limit is not None else 100,
}
for i, queue in enumerate(
self._get_paginated_list("/annotation-queues", params=params)
):
yield ls_schemas.AnnotationQueue(
**queue,
)
if limit is not None and i + 1 >= limit:
break
def create_annotation_queue(
self,
*,
name: str,
description: Optional[str] = None,
queue_id: Optional[ID_TYPE] = None,
) -> ls_schemas.AnnotationQueue:
"""Create an annotation queue on the LangSmith API.
Args:
name (str):
The name of the annotation queue.
description (Optional[str]):
The description of the annotation queue.
queue_id (Optional[Union[UUID, str]]):
The ID of the annotation queue.
Returns:
AnnotationQueue: The created annotation queue object.
"""
body = {
"name": name,
"description": description,
"id": queue_id or str(uuid.uuid4()),
}
response = self.request_with_retries(
"POST",
"/annotation-queues",
json={k: v for k, v in body.items() if v is not None},
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.AnnotationQueue(
**response.json(),
)
def read_annotation_queue(self, queue_id: ID_TYPE) -> ls_schemas.AnnotationQueue:
"""Read an annotation queue with the specified queue ID.
Args:
queue_id (Union[UUID, str]): The ID of the annotation queue to read.
Returns:
AnnotationQueue: The annotation queue object.
"""
# TODO: Replace when actual endpoint is added
return next(self.list_annotation_queues(queue_ids=[queue_id]))
def update_annotation_queue(
self, queue_id: ID_TYPE, *, name: str, description: Optional[str] = None
) -> None:
"""Update an annotation queue with the specified queue_id.
Args:
queue_id (Union[UUID, str]): The ID of the annotation queue to update.
name (str): The new name for the annotation queue.
description (Optional[str]): The new description for the
annotation queue. Defaults to None.
Returns:
None
"""
response = self.request_with_retries(
"PATCH",
f"/annotation-queues/{_as_uuid(queue_id, 'queue_id')}",
json={
"name": name,
"description": description,
},
)
ls_utils.raise_for_status_with_text(response)
def delete_annotation_queue(self, queue_id: ID_TYPE) -> None:
"""Delete an annotation queue with the specified queue ID.
Args:
queue_id (Union[UUID, str]): The ID of the annotation queue to delete.
Returns:
None
"""
response = self.request_with_retries(
"DELETE",
f"/annotation-queues/{_as_uuid(queue_id, 'queue_id')}",
headers={"Accept": "application/json", **self._headers},
)
ls_utils.raise_for_status_with_text(response)
def add_runs_to_annotation_queue(
self, queue_id: ID_TYPE, *, run_ids: List[ID_TYPE]
) -> None:
"""Add runs to an annotation queue with the specified queue ID.
Args:
queue_id (Union[UUID, str]): The ID of the annotation queue.
run_ids (List[Union[UUID, str]]): The IDs of the runs to be added to the annotation
queue.
Returns:
None
"""
response = self.request_with_retries(
"POST",
f"/annotation-queues/{_as_uuid(queue_id, 'queue_id')}/runs",
json=[str(_as_uuid(id_, f"run_ids[{i}]")) for i, id_ in enumerate(run_ids)],
)
ls_utils.raise_for_status_with_text(response)
def delete_run_from_annotation_queue(
self, queue_id: ID_TYPE, *, run_id: ID_TYPE
) -> None:
"""Delete a run from an annotation queue with the specified queue ID and run ID.
Args:
queue_id (Union[UUID, str]): The ID of the annotation queue.
run_id (Union[UUID, str]): The ID of the run to be added to the annotation
queue.
Returns:
None
"""
response = self.request_with_retries(
"DELETE",
f"/annotation-queues/{_as_uuid(queue_id, 'queue_id')}/runs/{_as_uuid(run_id, 'run_id')}",
)
ls_utils.raise_for_status_with_text(response)
def get_run_from_annotation_queue(
self, queue_id: ID_TYPE, *, index: int
) -> ls_schemas.RunWithAnnotationQueueInfo:
"""Get a run from an annotation queue at the specified index.
Args:
queue_id (Union[UUID, str]): The ID of the annotation queue.
index (int): The index of the run to retrieve.
Returns:
RunWithAnnotationQueueInfo: The run at the specified index.
Raises:
LangSmithNotFoundError: If the run is not found at the given index.
LangSmithError: For other API-related errors.
"""
base_url = f"/annotation-queues/{_as_uuid(queue_id, 'queue_id')}/run"
response = self.request_with_retries(
"GET",
f"{base_url}/{index}",
headers=self._headers,
)
ls_utils.raise_for_status_with_text(response)
return ls_schemas.RunWithAnnotationQueueInfo(**response.json())
def create_comparative_experiment(
self,
name: str,
experiments: Sequence[ID_TYPE],
*,
reference_dataset: Optional[ID_TYPE] = None,
description: Optional[str] = None,
created_at: Optional[datetime.datetime] = None,
metadata: Optional[Dict[str, Any]] = None,
id: Optional[ID_TYPE] = None,
) -> ls_schemas.ComparativeExperiment:
"""Create a comparative experiment on the LangSmith API.
These experiments compare 2 or more experiment results over a shared dataset.
Args:
name (str): The name of the comparative experiment.
experiments (Sequence[Union[UUID, str]]): The IDs of the experiments to compare.
reference_dataset (Optional[Union[UUID, str]]): The ID of the dataset these experiments are compared on.
description (Optional[str]): The description of the comparative experiment.
created_at (Optional[datetime.datetime]): The creation time of the comparative experiment.
metadata (Optional[Dict[str, Any]]): Additional metadata for the comparative experiment.
id (Optional[Union[UUID, str]]): The ID of the comparative experiment.
Returns:
ComparativeExperiment: The created comparative experiment object.
"""
if not experiments:
raise ValueError("At least one experiment is required.")
if reference_dataset is None:
# Get one of the experiments' reference dataset
reference_dataset = self.read_project(
project_id=experiments[0]
).reference_dataset_id
if not reference_dataset:
raise ValueError("A reference dataset is required.")
body: Dict[str, Any] = {
"id": id or str(uuid.uuid4()),
"name": name,
"experiment_ids": experiments,
"reference_dataset_id": reference_dataset,
"description": description,
"created_at": created_at or datetime.datetime.now(datetime.timezone.utc),
"extra": {},
}
if metadata is not None:
body["extra"]["metadata"] = metadata
ser = _dumps_json({k: v for k, v in body.items()}) # if v is not None})
response = self.request_with_retries(
"POST",
"/datasets/comparative",
request_kwargs={
"data": ser,
},
)
ls_utils.raise_for_status_with_text(response)
response_d = response.json()
return ls_schemas.ComparativeExperiment(**response_d)
async def arun_on_dataset(
self,
dataset_name: str,
llm_or_chain_factory: Any,
*,
evaluation: Optional[Any] = None,
concurrency_level: int = 5,
project_name: Optional[str] = None,
project_metadata: Optional[Dict[str, Any]] = None,
dataset_version: Optional[Union[datetime.datetime, str]] = None,
verbose: bool = False,
input_mapper: Optional[Callable[[Dict], Any]] = None,
revision_id: Optional[str] = None,
**kwargs: Any,
) -> Dict[str, Any]:
"""Asynchronously run the Chain or language model on a dataset.
.. deprecated:: 0.1.0
This method is deprecated. Use :func:`langsmith.aevaluate` instead.
""" # noqa: E501
warnings.warn(
"The `arun_on_dataset` method is deprecated and"
" will be removed in a future version."
"Please use the `aevaluate` method instead.",
DeprecationWarning,
)
try:
from langchain.smith import arun_on_dataset as _arun_on_dataset
except ImportError:
raise ImportError(
"The client.arun_on_dataset function requires the langchain"
"package to run.\nInstall with pip install langchain"
)
return await _arun_on_dataset(
dataset_name=dataset_name,
llm_or_chain_factory=llm_or_chain_factory,
client=self,
evaluation=evaluation,
concurrency_level=concurrency_level,
project_name=project_name,
project_metadata=project_metadata,
verbose=verbose,
input_mapper=input_mapper,
revision_id=revision_id,
dataset_version=dataset_version,
**kwargs,
)
def run_on_dataset(
self,
dataset_name: str,
llm_or_chain_factory: Any,
*,
evaluation: Optional[Any] = None,
concurrency_level: int = 5,
project_name: Optional[str] = None,
project_metadata: Optional[Dict[str, Any]] = None,
dataset_version: Optional[Union[datetime.datetime, str]] = None,
verbose: bool = False,
input_mapper: Optional[Callable[[Dict], Any]] = None,
revision_id: Optional[str] = None,
**kwargs: Any,
) -> Dict[str, Any]:
"""Run the Chain or language model on a dataset.
.. deprecated:: 0.1.0
This method is deprecated. Use :func:`langsmith.aevaluate` instead.
""" # noqa: E501 # noqa: E501
warnings.warn(
"The `run_on_dataset` method is deprecated and"
" will be removed in a future version."
"Please use the `evaluate` method instead.",
DeprecationWarning,
)
try:
from langchain.smith import (
run_on_dataset as _run_on_dataset, # type: ignore
)
except ImportError:
raise ImportError(
"The client.run_on_dataset function requires the langchain"
"package to run.\nInstall with pip install langchain"
)
return _run_on_dataset(
dataset_name=dataset_name,
llm_or_chain_factory=llm_or_chain_factory,
concurrency_level=concurrency_level,
client=self,
evaluation=evaluation,
project_name=project_name,
project_metadata=project_metadata,
verbose=verbose,
input_mapper=input_mapper,
revision_id=revision_id,
dataset_version=dataset_version,
**kwargs,
)
def _current_tenant_is_owner(self, owner: str) -> bool:
"""Check if the current workspace has the same handle as owner.
Args:
owner (str): The owner to check against.
Returns:
bool: True if the current tenant is the owner, False otherwise.
"""
settings = self._get_settings()
return owner == "-" or settings.tenant_handle == owner
def _owner_conflict_error(
self, action: str, owner: str
) -> ls_utils.LangSmithUserError:
return ls_utils.LangSmithUserError(
f"Cannot {action} for another tenant.\n"
f"Current tenant: {self._get_settings().tenant_handle},\n"
f"Requested tenant: {owner}"
)
def _get_latest_commit_hash(
self, prompt_owner_and_name: str, limit: int = 1, offset: int = 0
) -> Optional[str]:
"""Get the latest commit hash for a prompt.
Args:
prompt_owner_and_name (str): The owner and name of the prompt.
limit (int, default=1): The maximum number of commits to fetch. Defaults to 1.
offset (int, default=0): The number of commits to skip. Defaults to 0.
Returns:
Optional[str]: The latest commit hash, or None if no commits are found.
"""
response = self.request_with_retries(
"GET",
f"/commits/{prompt_owner_and_name}/",
params={"limit": limit, "offset": offset},
)
commits = response.json()["commits"]
return commits[0]["commit_hash"] if commits else None
def _like_or_unlike_prompt(
self, prompt_identifier: str, like: bool
) -> Dict[str, int]:
"""Like or unlike a prompt.
Args:
prompt_identifier (str): The identifier of the prompt.
like (bool): True to like the prompt, False to unlike it.
Returns:
A dictionary with the key 'likes' and the count of likes as the value.
Raises:
requests.exceptions.HTTPError: If the prompt is not found or
another error occurs.
"""
owner, prompt_name, _ = ls_utils.parse_prompt_identifier(prompt_identifier)
response = self.request_with_retries(
"POST", f"/likes/{owner}/{prompt_name}", json={"like": like}
)
response.raise_for_status()
return response.json()
def _get_prompt_url(self, prompt_identifier: str) -> str:
"""Get a URL for a prompt.
Args:
prompt_identifier (str): The identifier of the prompt.
Returns:
str: The URL for the prompt.
"""
owner, prompt_name, commit_hash = ls_utils.parse_prompt_identifier(
prompt_identifier
)
if not self._current_tenant_is_owner(owner):
return f"{self._host_url}/hub/{owner}/{prompt_name}:{commit_hash[:8]}"
settings = self._get_settings()
return (
f"{self._host_url}/prompts/{prompt_name}/{commit_hash[:8]}"
f"?organizationId={settings.id}"
)
def _prompt_exists(self, prompt_identifier: str) -> bool:
"""Check if a prompt exists.
Args:
prompt_identifier (str): The identifier of the prompt.
Returns:
bool: True if the prompt exists, False otherwise.
"""
prompt = self.get_prompt(prompt_identifier)
return True if prompt else False
def like_prompt(self, prompt_identifier: str) -> Dict[str, int]:
"""Like a prompt.
Args:
prompt_identifier (str): The identifier of the prompt.
Returns:
Dict[str, int]: A dictionary with the key 'likes' and the count of likes as the value.
"""
return self._like_or_unlike_prompt(prompt_identifier, like=True)
def unlike_prompt(self, prompt_identifier: str) -> Dict[str, int]:
"""Unlike a prompt.
Args:
prompt_identifier (str): The identifier of the prompt.
Returns:
Dict[str, int]: A dictionary with the key 'likes' and the count of likes as the value.
"""
return self._like_or_unlike_prompt(prompt_identifier, like=False)
def list_prompts(
self,
*,
limit: int = 100,
offset: int = 0,
is_public: Optional[bool] = None,
is_archived: Optional[bool] = False,
sort_field: ls_schemas.PromptSortField = ls_schemas.PromptSortField.updated_at,
sort_direction: Literal["desc", "asc"] = "desc",
query: Optional[str] = None,
) -> ls_schemas.ListPromptsResponse:
"""List prompts with pagination.
Args:
limit (int, default=100): The maximum number of prompts to return. Defaults to 100.
offset (int, default=0): The number of prompts to skip. Defaults to 0.
is_public (Optional[bool]): Filter prompts by if they are public.
is_archived (Optional[bool]): Filter prompts by if they are archived.
sort_field (PromptSortField): The field to sort by.
Defaults to "updated_at".
sort_direction (Literal["desc", "asc"], default="desc"): The order to sort by.
Defaults to "desc".
query (Optional[str]): Filter prompts by a search query.
Returns:
ListPromptsResponse: A response object containing
the list of prompts.
"""
params = {
"limit": limit,
"offset": offset,
"is_public": (
"true" if is_public else "false" if is_public is not None else None
),
"is_archived": "true" if is_archived else "false",
"sort_field": sort_field,
"sort_direction": sort_direction,
"query": query,
"match_prefix": "true" if query else None,
}
response = self.request_with_retries("GET", "/repos/", params=params)
return ls_schemas.ListPromptsResponse(**response.json())
def get_prompt(self, prompt_identifier: str) -> Optional[ls_schemas.Prompt]:
"""Get a specific prompt by its identifier.
Args:
prompt_identifier (str): The identifier of the prompt.
The identifier should be in the format "prompt_name" or "owner/prompt_name".
Returns:
Optional[Prompt]: The prompt object.
Raises:
requests.exceptions.HTTPError: If the prompt is not found or
another error occurs.
"""
owner, prompt_name, _ = ls_utils.parse_prompt_identifier(prompt_identifier)
try:
response = self.request_with_retries("GET", f"/repos/{owner}/{prompt_name}")
return ls_schemas.Prompt(**response.json()["repo"])
except ls_utils.LangSmithNotFoundError:
return None
def create_prompt(
self,
prompt_identifier: str,
*,
description: Optional[str] = None,
readme: Optional[str] = None,
tags: Optional[Sequence[str]] = None,
is_public: bool = False,
) -> ls_schemas.Prompt:
"""Create a new prompt.
Does not attach prompt object, just creates an empty prompt.
Args:
prompt_identifier (str): The identifier of the prompt.
The identifier should be in the formatof owner/name:hash, name:hash, owner/name, or name
description (Optional[str]): A description of the prompt.
readme (Optional[str]): A readme for the prompt.
tags (Optional[Sequence[str]]): A list of tags for the prompt.
is_public (bool): Whether the prompt should be public. Defaults to False.
Returns:
Prompt: The created prompt object.
Raises:
ValueError: If the current tenant is not the owner.
HTTPError: If the server request fails.
"""
settings = self._get_settings()
if is_public and not settings.tenant_handle:
raise ls_utils.LangSmithUserError(
"Cannot create a public prompt without first\n"
"creating a LangChain Hub handle. "
"You can add a handle by creating a public prompt at:\n"
"https://smith.langchain.com/prompts"
)
owner, prompt_name, _ = ls_utils.parse_prompt_identifier(prompt_identifier)
if not self._current_tenant_is_owner(owner=owner):
raise self._owner_conflict_error("create a prompt", owner)
json: Dict[str, Union[str, bool, Sequence[str]]] = {
"repo_handle": prompt_name,
"description": description or "",
"readme": readme or "",
"tags": tags or [],
"is_public": is_public,
}
response = self.request_with_retries("POST", "/repos/", json=json)
response.raise_for_status()
return ls_schemas.Prompt(**response.json()["repo"])
def create_commit(
self,
prompt_identifier: str,
object: Any,
*,
parent_commit_hash: Optional[str] = None,
) -> str:
"""Create a commit for an existing prompt.
Args:
prompt_identifier (str): The identifier of the prompt.
object (Any): The LangChain object to commit.
parent_commit_hash (Optional[str]): The hash of the parent commit.
Defaults to latest commit.
Returns:
str: The url of the prompt commit.
Raises:
HTTPError: If the server request fails.
ValueError: If the prompt does not exist.
"""
if not self._prompt_exists(prompt_identifier):
raise ls_utils.LangSmithNotFoundError(
"Prompt does not exist, you must create it first."
)
try:
from langchain_core.load.dump import dumps
except ImportError:
raise ImportError(
"The client.create_commit function requires the langchain_core"
"package to run.\nInstall with `pip install langchain_core`"
)
json_object = dumps(object)
manifest_dict = json.loads(json_object)
owner, prompt_name, _ = ls_utils.parse_prompt_identifier(prompt_identifier)
prompt_owner_and_name = f"{owner}/{prompt_name}"
if parent_commit_hash == "latest" or parent_commit_hash is None:
parent_commit_hash = self._get_latest_commit_hash(prompt_owner_and_name)
request_dict = {"parent_commit": parent_commit_hash, "manifest": manifest_dict}
response = self.request_with_retries(
"POST", f"/commits/{prompt_owner_and_name}", json=request_dict
)
commit_hash = response.json()["commit"]["commit_hash"]
return self._get_prompt_url(f"{prompt_owner_and_name}:{commit_hash}")
def update_prompt(
self,
prompt_identifier: str,
*,
description: Optional[str] = None,
readme: Optional[str] = None,
tags: Optional[Sequence[str]] = None,
is_public: Optional[bool] = None,
is_archived: Optional[bool] = None,
) -> Dict[str, Any]:
"""Update a prompt's metadata.
To update the content of a prompt, use push_prompt or create_commit instead.
Args:
prompt_identifier (str): The identifier of the prompt to update.
description (Optional[str]): New description for the prompt.
readme (Optional[str]): New readme for the prompt.
tags (Optional[Sequence[str]]): New list of tags for the prompt.
is_public (Optional[bool]): New public status for the prompt.
is_archived (Optional[bool]): New archived status for the prompt.
Returns:
Dict[str, Any]: The updated prompt data as returned by the server.
Raises:
ValueError: If the prompt_identifier is empty.
HTTPError: If the server request fails.
"""
settings = self._get_settings()
if is_public and not settings.tenant_handle:
raise ValueError(
"Cannot create a public prompt without first\n"
"creating a LangChain Hub handle. "
"You can add a handle by creating a public prompt at:\n"
"https://smith.langchain.com/prompts"
)
json: Dict[str, Union[str, bool, Sequence[str]]] = {}
if description is not None:
json["description"] = description
if readme is not None:
json["readme"] = readme
if is_public is not None:
json["is_public"] = is_public
if is_archived is not None:
json["is_archived"] = is_archived
if tags is not None:
json["tags"] = tags
owner, prompt_name, _ = ls_utils.parse_prompt_identifier(prompt_identifier)
response = self.request_with_retries(
"PATCH", f"/repos/{owner}/{prompt_name}", json=json
)
response.raise_for_status()
return response.json()
def delete_prompt(self, prompt_identifier: str) -> None:
"""Delete a prompt.
Args:
prompt_identifier (str): The identifier of the prompt to delete.
Returns:
bool: True if the prompt was successfully deleted, False otherwise.
Raises:
ValueError: If the current tenant is not the owner of the prompt.
"""
owner, prompt_name, _ = ls_utils.parse_prompt_identifier(prompt_identifier)
if not self._current_tenant_is_owner(owner):
raise self._owner_conflict_error("delete a prompt", owner)
response = self.request_with_retries("DELETE", f"/repos/{owner}/{prompt_name}")
response.raise_for_status()
def pull_prompt_commit(
self,
prompt_identifier: str,
*,
include_model: Optional[bool] = False,
) -> ls_schemas.PromptCommit:
"""Pull a prompt object from the LangSmith API.
Args:
prompt_identifier (str): The identifier of the prompt.
Returns:
PromptCommit: The prompt object.
Raises:
ValueError: If no commits are found for the prompt.
"""
owner, prompt_name, commit_hash = ls_utils.parse_prompt_identifier(
prompt_identifier
)
response = self.request_with_retries(
"GET",
(
f"/commits/{owner}/{prompt_name}/{commit_hash}"
f"{'?include_model=true' if include_model else ''}"
),
)
return ls_schemas.PromptCommit(
**{"owner": owner, "repo": prompt_name, **response.json()}
)
def list_prompt_commits(
self,
prompt_identifier: str,
*,
limit: Optional[int] = None,
offset: int = 0,
include_model: bool = False,
) -> Iterator[ls_schemas.ListedPromptCommit]:
"""List commits for a given prompt.
Args:
prompt_identifier (str): The identifier of the prompt in the format 'owner/repo_name'.
limit (Optional[int]): The maximum number of commits to return. If None, returns all commits. Defaults to None.
offset (int, default=0): The number of commits to skip before starting to return results. Defaults to 0.
include_model (bool, default=False): Whether to include the model information in the commit data. Defaults to False.
Yields:
A ListedPromptCommit object for each commit.
Note:
This method uses pagination to retrieve commits. It will make multiple API calls if necessary to retrieve all commits
or up to the specified limit.
"""
owner, prompt_name, _ = ls_utils.parse_prompt_identifier(prompt_identifier)
params = {
"limit": min(100, limit) if limit is not None else limit,
"offset": offset,
"include_model": include_model,
}
i = 0
while True:
params["offset"] = offset
response = self.request_with_retries(
"GET",
f"/commits/{owner}/{prompt_name}/",
params=params,
)
val = response.json()
items = val["commits"]
total = val["total"]
if not items:
break
for it in items:
if limit is not None and i >= limit:
return # Stop iteration if we've reached the limit
yield ls_schemas.ListedPromptCommit(
**{"owner": owner, "repo": prompt_name, **it}
)
i += 1
offset += len(items)
if offset >= total:
break
def pull_prompt(
self, prompt_identifier: str, *, include_model: Optional[bool] = False
) -> Any:
"""Pull a prompt and return it as a LangChain PromptTemplate.
This method requires `langchain_core`.
Args:
prompt_identifier (str): The identifier of the prompt.
include_model (Optional[bool], default=False): Whether to include the model information in the prompt data.
Returns:
Any: The prompt object in the specified format.
"""
try:
from langchain_core.language_models.base import BaseLanguageModel
from langchain_core.load.load import loads
from langchain_core.output_parsers import BaseOutputParser
from langchain_core.prompts import BasePromptTemplate
from langchain_core.prompts.structured import StructuredPrompt
from langchain_core.runnables.base import RunnableBinding, RunnableSequence
except ImportError:
raise ImportError(
"The client.pull_prompt function requires the langchain_core"
"package to run.\nInstall with `pip install langchain_core`"
)
try:
from langchain_core._api import suppress_langchain_beta_warning
except ImportError:
@contextlib.contextmanager
def suppress_langchain_beta_warning():
yield
prompt_object = self.pull_prompt_commit(
prompt_identifier, include_model=include_model
)
with suppress_langchain_beta_warning():
prompt = loads(json.dumps(prompt_object.manifest))
if (
isinstance(prompt, BasePromptTemplate)
or isinstance(prompt, RunnableSequence)
and isinstance(prompt.first, BasePromptTemplate)
):
prompt_template = (
prompt
if isinstance(prompt, BasePromptTemplate)
else (
prompt.first
if isinstance(prompt, RunnableSequence)
and isinstance(prompt.first, BasePromptTemplate)
else None
)
)
if prompt_template is None:
raise ls_utils.LangSmithError(
"Prompt object is not a valid prompt template."
)
if prompt_template.metadata is None:
prompt_template.metadata = {}
prompt_template.metadata.update(
{
"lc_hub_owner": prompt_object.owner,
"lc_hub_repo": prompt_object.repo,
"lc_hub_commit_hash": prompt_object.commit_hash,
}
)
if (
include_model
and isinstance(prompt, RunnableSequence)
and isinstance(prompt.first, StructuredPrompt)
# Make forward-compatible in case we let update the response type
and (
len(prompt.steps) == 2 and not isinstance(prompt.last, BaseOutputParser)
)
):
if isinstance(prompt.last, RunnableBinding) and isinstance(
prompt.last.bound, BaseLanguageModel
):
seq = cast(RunnableSequence, prompt.first | prompt.last.bound)
if len(seq.steps) == 3: # prompt | bound llm | output parser
rebound_llm = seq.steps[1]
prompt = RunnableSequence(
prompt.first,
rebound_llm.bind(**{**prompt.last.kwargs}),
seq.last,
)
else:
prompt = seq # Not sure
elif isinstance(prompt.last, BaseLanguageModel):
prompt: RunnableSequence = prompt.first | prompt.last # type: ignore[no-redef, assignment]
else:
pass
return prompt
def push_prompt(
self,
prompt_identifier: str,
*,
object: Optional[Any] = None,
parent_commit_hash: str = "latest",
is_public: Optional[bool] = None,
description: Optional[str] = None,
readme: Optional[str] = None,
tags: Optional[Sequence[str]] = None,
) -> str:
"""Push a prompt to the LangSmith API.
Can be used to update prompt metadata or prompt content.
If the prompt does not exist, it will be created.
If the prompt exists, it will be updated.
Args:
prompt_identifier (str): The identifier of the prompt.
object (Optional[Any]): The LangChain object to push.
parent_commit_hash (str): The parent commit hash.
Defaults to "latest".
is_public (Optional[bool]): Whether the prompt should be public.
If None (default), the current visibility status is maintained for existing prompts.
For new prompts, None defaults to private.
Set to True to make public, or False to make private.
description (Optional[str]): A description of the prompt.
Defaults to an empty string.
readme (Optional[str]): A readme for the prompt.
Defaults to an empty string.
tags (Optional[Sequence[str]]): A list of tags for the prompt.
Defaults to an empty list.
Returns:
str: The URL of the prompt.
"""
# Create or update prompt metadata
if self._prompt_exists(prompt_identifier):
if any(
param is not None for param in [is_public, description, readme, tags]
):
self.update_prompt(
prompt_identifier,
description=description,
readme=readme,
tags=tags,
is_public=is_public,
)
else:
self.create_prompt(
prompt_identifier,
is_public=is_public if is_public is not None else False,
description=description,
readme=readme,
tags=tags,
)
if object is None:
return self._get_prompt_url(prompt_identifier=prompt_identifier)
# Create a commit with the new manifest
url = self.create_commit(
prompt_identifier,
object,
parent_commit_hash=parent_commit_hash,
)
return url
def cleanup(self) -> None:
"""Manually trigger cleanup of the background thread."""
self._manual_cleanup = True
@overload
def evaluate(
self,
target: Union[TARGET_T, Runnable, EXPERIMENT_T],
/,
data: Optional[DATA_T] = None,
evaluators: Optional[Sequence[EVALUATOR_T]] = None,
summary_evaluators: Optional[Sequence[SUMMARY_EVALUATOR_T]] = None,
metadata: Optional[dict] = None,
experiment_prefix: Optional[str] = None,
description: Optional[str] = None,
max_concurrency: Optional[int] = 0,
num_repetitions: int = 1,
blocking: bool = True,
experiment: Optional[EXPERIMENT_T] = None,
upload_results: bool = True,
**kwargs: Any,
) -> ExperimentResults: ...
@overload
def evaluate(
self,
target: Union[Tuple[EXPERIMENT_T, EXPERIMENT_T]],
/,
data: Optional[DATA_T] = None,
evaluators: Optional[Sequence[COMPARATIVE_EVALUATOR_T]] = None,
summary_evaluators: Optional[Sequence[SUMMARY_EVALUATOR_T]] = None,
metadata: Optional[dict] = None,
experiment_prefix: Optional[str] = None,
description: Optional[str] = None,
max_concurrency: Optional[int] = 0,
num_repetitions: int = 1,
blocking: bool = True,
experiment: Optional[EXPERIMENT_T] = None,
upload_results: bool = True,
**kwargs: Any,
) -> ComparativeExperimentResults: ...
def evaluate(
self,
target: Union[
TARGET_T, Runnable, EXPERIMENT_T, Tuple[EXPERIMENT_T, EXPERIMENT_T]
],
/,
data: Optional[DATA_T] = None,
evaluators: Optional[
Union[Sequence[EVALUATOR_T], Sequence[COMPARATIVE_EVALUATOR_T]]
] = None,
summary_evaluators: Optional[Sequence[SUMMARY_EVALUATOR_T]] = None,
metadata: Optional[dict] = None,
experiment_prefix: Optional[str] = None,
description: Optional[str] = None,
max_concurrency: Optional[int] = 0,
num_repetitions: int = 1,
blocking: bool = True,
experiment: Optional[EXPERIMENT_T] = None,
upload_results: bool = True,
**kwargs: Any,
) -> Union[ExperimentResults, ComparativeExperimentResults]:
r"""Evaluate a target system on a given dataset.
Args:
target (Union[TARGET_T, Runnable, EXPERIMENT_T, Tuple[EXPERIMENT_T, EXPERIMENT_T]]):
The target system or experiment(s) to evaluate. Can be a function
that takes a dict and returns a dict, a langchain Runnable, an
existing experiment ID, or a two-tuple of experiment IDs.
data (DATA_T): The dataset to evaluate on. Can be a dataset name, a list of
examples, or a generator of examples.
evaluators (Optional[Union[Sequence[EVALUATOR_T], Sequence[COMPARATIVE_EVALUATOR_T]]]):
A list of evaluators to run on each example. The evaluator signature
depends on the target type. Default to None.
summary_evaluators (Optional[Sequence[SUMMARY_EVALUATOR_T]]): A list of summary
evaluators to run on the entire dataset. Should not be specified if
comparing two existing experiments. Defaults to None.
metadata (Optional[dict]): Metadata to attach to the experiment.
Defaults to None.
experiment_prefix (Optional[str]): A prefix to provide for your experiment name.
Defaults to None.
description (Optional[str]): A free-form text description for the experiment.
max_concurrency (Optional[int], default=0): The maximum number of concurrent
evaluations to run. If None then no limit is set. If 0 then no concurrency.
Defaults to 0.
blocking (bool, default=True): Whether to block until the evaluation is complete.
Defaults to True.
num_repetitions (int, default=1): The number of times to run the evaluation.
Each item in the dataset will be run and evaluated this many times.
Defaults to 1.
experiment (Optional[EXPERIMENT_T]): An existing experiment to
extend. If provided, experiment_prefix is ignored. For advanced
usage only. Should not be specified if target is an existing experiment or
two-tuple fo experiments.
upload_results (bool, default=True): Whether to upload the results to LangSmith.
Defaults to True.
**kwargs (Any): Additional keyword arguments to pass to the evaluator.
Returns:
ExperimentResults: If target is a function, Runnable, or existing experiment.
ComparativeExperimentResults: If target is a two-tuple of existing experiments.
Examples:
Prepare the dataset:
.. code-block:: python
from langsmith import Client
client = Client()
dataset = client.clone_public_dataset(
"https://smith.langchain.com/public/419dcab2-1d66-4b94-8901-0357ead390df/d"
)
dataset_name = "Evaluate Examples"
Basic usage:
.. code-block:: python
def accuracy(outputs: dict, reference_outputs: dict) -> dict:
# Row-level evaluator for accuracy.
pred = outputs["response"]
expected = reference_outputs["answer"]
return {"score": expected.lower() == pred.lower()}
.. code-block:: python
def precision(outputs: list[dict], reference_outputs: list[dict]) -> dict:
# Experiment-level evaluator for precision.
# TP / (TP + FP)
predictions = [out["response"].lower() for out in outputs]
expected = [ref["answer"].lower() for ref in reference_outputs]
# yes and no are the only possible answers
tp = sum([p == e for p, e in zip(predictions, expected) if p == "yes"])
fp = sum([p == "yes" and e == "no" for p, e in zip(predictions, expected)])
return {"score": tp / (tp + fp)}
def predict(inputs: dict) -> dict:
# This can be any function or just an API call to your app.
return {"response": "Yes"}
results = client.evaluate(
predict,
data=dataset_name,
evaluators=[accuracy],
summary_evaluators=[precision],
experiment_prefix="My Experiment",
description="Evaluating the accuracy of a simple prediction model.",
metadata={
"my-prompt-version": "abcd-1234",
},
)
Evaluating over only a subset of the examples
.. code-block:: python
experiment_name = results.experiment_name
examples = client.list_examples(dataset_name=dataset_name, limit=5)
results = client.evaluate(
predict,
data=examples,
evaluators=[accuracy],
summary_evaluators=[precision],
experiment_prefix="My Experiment",
description="Just testing a subset synchronously.",
)
Streaming each prediction to more easily + eagerly debug.
.. code-block:: python
results = client.evaluate(
predict,
data=dataset_name,
evaluators=[accuracy],
summary_evaluators=[precision],
description="I don't even have to block!",
blocking=False,
)
for i, result in enumerate(results): # doctest: +ELLIPSIS
pass
Using the `evaluate` API with an off-the-shelf LangChain evaluator:
.. code-block:: python
from langsmith.evaluation import LangChainStringEvaluator
from langchain.chat_models import init_chat_model
def prepare_criteria_data(run: Run, example: Example):
return {
"prediction": run.outputs["output"],
"reference": example.outputs["answer"],
"input": str(example.inputs),
}
results = client.evaluate(
predict,
data=dataset_name,
evaluators=[
accuracy,
LangChainStringEvaluator("embedding_distance"),
LangChainStringEvaluator(
"labeled_criteria",
config={
"criteria": {
"usefulness": "The prediction is useful if it is correct"
" and/or asks a useful followup question."
},
"llm": init_chat_model("gpt-4o"),
},
prepare_data=prepare_criteria_data,
),
],
description="Evaluating with off-the-shelf LangChain evaluators.",
summary_evaluators=[precision],
)
View the evaluation results for experiment:...
Evaluating a LangChain object:
.. code-block:: python
from langchain_core.runnables import chain as as_runnable
@as_runnable
def nested_predict(inputs):
return {"response": "Yes"}
@as_runnable
def lc_predict(inputs):
return nested_predict.invoke(inputs)
results = client.evaluate(
lc_predict,
data=dataset_name,
evaluators=[accuracy],
description="This time we're evaluating a LangChain object.",
summary_evaluators=[precision],
)
Comparative evaluation:
.. code-block:: python
results = client.evaluate(
# The target is a tuple of the experiment IDs to compare
target=(
"12345678-1234-1234-1234-123456789012",
"98765432-1234-1234-1234-123456789012",
),
evaluators=[accuracy],
summary_evaluators=[precision],
)
Evaluate an existing experiment:
.. code-block:: python
results = client.evaluate(
# The target is the ID of the experiment we are evaluating
target="12345678-1234-1234-1234-123456789012",
evaluators=[accuracy],
summary_evaluators=[precision],
)
.. versionadded:: 0.2.0
""" # noqa: E501
from langsmith.evaluation._runner import evaluate as evaluate_
# Need to ignore because it fails when there are too many union types +
# overloads.
return evaluate_( # type: ignore[misc]
target, # type: ignore[arg-type]
data=data,
evaluators=evaluators, # type: ignore[arg-type]
summary_evaluators=summary_evaluators,
metadata=metadata,
experiment_prefix=experiment_prefix,
description=description,
max_concurrency=max_concurrency,
num_repetitions=num_repetitions,
client=self,
blocking=blocking,
experiment=experiment,
upload_results=upload_results,
**kwargs,
)
async def aevaluate(
self,
target: Union[
ATARGET_T,
AsyncIterable[dict],
Runnable,
str,
uuid.UUID,
schemas.TracerSession,
],
/,
data: Union[
DATA_T, AsyncIterable[schemas.Example], Iterable[schemas.Example], None
] = None,
evaluators: Optional[Sequence[Union[EVALUATOR_T, AEVALUATOR_T]]] = None,
summary_evaluators: Optional[Sequence[SUMMARY_EVALUATOR_T]] = None,
metadata: Optional[dict] = None,
experiment_prefix: Optional[str] = None,
description: Optional[str] = None,
max_concurrency: Optional[int] = 0,
num_repetitions: int = 1,
blocking: bool = True,
experiment: Optional[Union[schemas.TracerSession, str, uuid.UUID]] = None,
upload_results: bool = True,
**kwargs: Any,
) -> AsyncExperimentResults:
r"""Evaluate an async target system on a given dataset.
Args:
target (Union[ATARGET_T, AsyncIterable[dict], Runnable, str, uuid.UUID, TracerSession]):
The target system or experiment(s) to evaluate. Can be an async function
that takes a dict and returns a dict, a langchain Runnable, an
existing experiment ID, or a two-tuple of experiment IDs.
data (Union[DATA_T, AsyncIterable[Example]]): The dataset to evaluate on. Can be a dataset name, a list of
examples, an async generator of examples, or an async iterable of examples.
evaluators (Optional[Sequence[EVALUATOR_T]]): A list of evaluators to run
on each example. Defaults to None.
summary_evaluators (Optional[Sequence[SUMMARY_EVALUATOR_T]]): A list of summary
evaluators to run on the entire dataset. Defaults to None.
metadata (Optional[dict]): Metadata to attach to the experiment.
Defaults to None.
experiment_prefix (Optional[str]): A prefix to provide for your experiment name.
Defaults to None.
description (Optional[str]): A description of the experiment.
max_concurrency (Optional[int], default=0): The maximum number of concurrent
evaluations to run. If None then no limit is set. If 0 then no concurrency.
Defaults to 0.
num_repetitions (int, default=1): The number of times to run the evaluation.
Each item in the dataset will be run and evaluated this many times.
Defaults to 1.
blocking (bool, default=True): Whether to block until the evaluation is complete.
Defaults to True.
experiment (Optional[TracerSession]): An existing experiment to
extend. If provided, experiment_prefix is ignored. For advanced
usage only.
upload_results (bool, default=True): Whether to upload the results to LangSmith.
Defaults to True.
**kwargs (Any): Additional keyword arguments to pass to the evaluator.
Returns:
AsyncIterator[ExperimentResultRow]: An async iterator over the experiment results.
Environment:
- LANGSMITH_TEST_CACHE: If set, API calls will be cached to disk to save time and
cost during testing. Recommended to commit the cache files to your repository
for faster CI/CD runs.
Requires the 'langsmith[vcr]' package to be installed.
Examples:
Prepare the dataset:
.. code-block:: python
import asyncio
from langsmith import Client
client = Client()
dataset = client.clone_public_dataset(
"https://smith.langchain.com/public/419dcab2-1d66-4b94-8901-0357ead390df/d"
)
dataset_name = "Evaluate Examples"
Basic usage:
.. code-block:: python
def accuracy(outputs: dict, reference_outputs: dict) -> dict:
# Row-level evaluator for accuracy.
pred = outputs["resposen"]
expected = reference_outputs["answer"]
return {"score": expected.lower() == pred.lower()}
def precision(outputs: list[dict], reference_outputs: list[dict]) -> dict:
# Experiment-level evaluator for precision.
# TP / (TP + FP)
predictions = [out["response"].lower() for out in outputs]
expected = [ref["answer"].lower() for ref in reference_outputs]
# yes and no are the only possible answers
tp = sum([p == e for p, e in zip(predictions, expected) if p == "yes"])
fp = sum([p == "yes" and e == "no" for p, e in zip(predictions, expected)])
return {"score": tp / (tp + fp)}
async def apredict(inputs: dict) -> dict:
# This can be any async function or just an API call to your app.
await asyncio.sleep(0.1)
return {"response": "Yes"}
results = asyncio.run(
client.aevaluate(
apredict,
data=dataset_name,
evaluators=[accuracy],
summary_evaluators=[precision],
experiment_prefix="My Experiment",
description="Evaluate the accuracy of the model asynchronously.",
metadata={
"my-prompt-version": "abcd-1234",
},
)
)
Evaluating over only a subset of the examples using an async generator:
.. code-block:: python
async def example_generator():
examples = client.list_examples(dataset_name=dataset_name, limit=5)
for example in examples:
yield example
results = asyncio.run(
client.aevaluate(
apredict,
data=example_generator(),
evaluators=[accuracy],
summary_evaluators=[precision],
experiment_prefix="My Subset Experiment",
description="Evaluate a subset of examples asynchronously.",
)
)
Streaming each prediction to more easily + eagerly debug.
.. code-block:: python
results = asyncio.run(
client.aevaluate(
apredict,
data=dataset_name,
evaluators=[accuracy],
summary_evaluators=[precision],
experiment_prefix="My Streaming Experiment",
description="Streaming predictions for debugging.",
blocking=False,
)
)
async def aenumerate(iterable):
async for elem in iterable:
print(elem)
asyncio.run(aenumerate(results))
Running without concurrency:
.. code-block:: python
results = asyncio.run(
client.aevaluate(
apredict,
data=dataset_name,
evaluators=[accuracy],
summary_evaluators=[precision],
experiment_prefix="My Experiment Without Concurrency",
description="This was run without concurrency.",
max_concurrency=0,
)
)
Using Async evaluators:
.. code-block:: python
async def helpfulness(outputs: dict) -> dict:
# Row-level evaluator for helpfulness.
await asyncio.sleep(5) # Replace with your LLM API call
return {"score": outputs["output"] == "Yes"}
results = asyncio.run(
client.aevaluate(
apredict,
data=dataset_name,
evaluators=[helpfulness],
summary_evaluators=[precision],
experiment_prefix="My Helpful Experiment",
description="Applying async evaluators example.",
)
)
Evaluate an existing experiment:
.. code-block:: python
results = asyncio.run(
client.aevaluate(
# The target is the ID of the experiment we are evaluating
target="419dcab2-1d66-4b94-8901-0357ead390df",
evaluators=[accuracy, helpfulness],
summary_evaluators=[precision],
)
)
.. versionadded:: 0.2.0
""" # noqa: E501
from langsmith.evaluation._arunner import aevaluate as aevaluate_
return await aevaluate_(
target,
data=data,
evaluators=evaluators,
summary_evaluators=summary_evaluators,
metadata=metadata,
experiment_prefix=experiment_prefix,
description=description,
max_concurrency=max_concurrency,
num_repetitions=num_repetitions,
client=self,
blocking=blocking,
experiment=experiment,
upload_results=upload_results,
**kwargs,
)
def convert_prompt_to_openai_format(
messages: Any,
model_kwargs: Optional[Dict[str, Any]] = None,
) -> dict:
"""Convert a prompt to OpenAI format.
Requires the `langchain_openai` package to be installed.
Args:
messages (Any): The messages to convert.
model_kwargs (Optional[Dict[str, Any]]): Model configuration arguments including
`stop` and any other required arguments. Defaults to None.
Returns:
dict: The prompt in OpenAI format.
Raises:
ImportError: If the `langchain_openai` package is not installed.
ls_utils.LangSmithError: If there is an error during the conversion process.
"""
try:
from langchain_openai import ChatOpenAI # type: ignore
except ImportError:
raise ImportError(
"The convert_prompt_to_openai_format function requires the langchain_openai"
"package to run.\nInstall with `pip install langchain_openai`"
)
openai = ChatOpenAI()
model_kwargs = model_kwargs or {}
stop = model_kwargs.pop("stop", None)
try:
return openai._get_request_payload(messages, stop=stop, **model_kwargs)
except Exception as e:
raise ls_utils.LangSmithError(f"Error converting to OpenAI format: {e}")
def convert_prompt_to_anthropic_format(
messages: Any,
model_kwargs: Optional[Dict[str, Any]] = None,
) -> dict:
"""Convert a prompt to Anthropic format.
Requires the `langchain_anthropic` package to be installed.
Args:
messages (Any): The messages to convert.
model_kwargs (Optional[Dict[str, Any]]):
Model configuration arguments including `model_name` and `stop`.
Defaults to None.
Returns:
dict: The prompt in Anthropic format.
"""
try:
from langchain_anthropic import ChatAnthropic # type: ignore
except ImportError:
raise ImportError(
"The convert_prompt_to_anthropic_format function requires the "
"langchain_anthropic package to run.\n"
"Install with `pip install langchain_anthropic`"
)
model_kwargs = model_kwargs or {}
model_name = model_kwargs.pop("model_name", "claude-3-haiku-20240307")
stop = model_kwargs.pop("stop", None)
timeout = model_kwargs.pop("timeout", None)
anthropic = ChatAnthropic(
model_name=model_name, timeout=timeout, stop=stop, **model_kwargs
)
try:
return anthropic._get_request_payload(messages, stop=stop)
except Exception as e:
raise ls_utils.LangSmithError(f"Error converting to Anthropic format: {e}")