Langchain

This integration connects Sentry with Langchain. The integration has been confirmed to work with Langchain 0.1.11.

Install

Install sentry-sdk from PyPI and the appropriate langchain packages:

pip install --upgrade 'sentry-sdk' 'langchain-openai' 'langchain-core'

Configure

If you have the langchain package in your dependencies, the Langchain integration will be enabled automatically when you initialize the Sentry SDK.

An additional dependency, tiktoken, is required to be installed if you want to calculate token usage for streaming chat responses.

import sentry_sdk

sentry_sdk.init(
    dsn="https://examplePublicKey@o0.ingest.sentry.io/0",
    send_default_pii=True, # send personally-identifiable information like LLM responses to sentry
    # Set traces_sample_rate to 1.0 to capture 100%
    # of transactions for performance monitoring.
    traces_sample_rate=1.0,
    # Set profiles_sample_rate to 1.0 to profile 100%
    # of sampled transactions.
    # We recommend adjusting this value in production.
    profiles_sample_rate=1.0,
)

Verify

Verify that the integration works by inducing an error:

from langchain_openai import ChatOpenAI
import sentry_sdk

sentry_sdk.init(...)  # same as above

llm = ChatOpenAI(model="gpt-3.5-turbo-0125", temperature=0, api_key="bad API key")
with sentry_sdk.start_transaction(op="ai-inference", name="The result of the AI inference"):
    response = llm.invoke([("system", "What is the capital of paris?")])
    print(response)

After running this script, a transaction will be created in the Performance section of sentry.io. Additionally, an error event (about the bad API key) will be sent to sentry.io and will be connected to the transaction.

It may take a couple of moments for the data to appear in sentry.io.

Behavior

• The Langchain integration will connect Sentry with Langchain and automatically monitor all LLM, tool, and function calls.

• All exceptions in the execution of the chain are reported.

• Sentry considers LLM and tokenizer inputs/outputs as PII and, by default, does not include PII data. If you want to include the data, set send_default_pii=True in the sentry_sdk.init() call. To explicitly exclude prompts and outputs despite send_default_pii=True, configure the integration with include_prompts=False as shown in the Options section below.

Options

By adding LangchainIntegration to your sentry_sdk.init() call explicitly, you can set options for LangchainIntegration to change its behavior:

import sentry_sdk
from sentry_sdk.integrations.langchain import LangchainIntegration

sentry_sdk.init(
    # ...
    send_default_pii=True,
    integrations = [
        LangchainIntegration(
            include_prompts=False,  # LLM/tokenizer inputs/outputs will be not sent to Sentry, despite send_default_pii=True
            max_spans=500,
            tiktoken_encoding_name="cl100k_base",
        ),
    ],
)

You can pass the following keyword arguments to LangchainIntegration():

• include_prompts:

  Whether LLM and tokenizer inputs and outputs should be sent to Sentry. Sentry considers this data personal identifiable data (PII) by default. If you want to include the data, set send_default_pii=True in the sentry_sdk.init() call. To explicitly exclude prompts and outputs despite send_default_pii=True, configure the integration with include_prompts=False.

  The default is True.

• max_spans:

  The most number of spans (e.g., LLM calls) that can be processed at the same time.

  The default is 1024.

• tiktoken_encoding_name:

  If you want to calculate token usage for streaming chat responses you need to have an additional dependency, tiktoken installed and specify the tiktoken_encoding_name that you use for tokenization. See the OpenAI Cookbook for possible values.

  The default is None.

Supported Versions

• Langchain: 0.1.11+
• tiktoken: 0.6.0+
• Python: 3.9+