LangChain Anthropic

Common errors

• anthropic.AuthenticationError: Invalid x-api-key.

  cause The Anthropic API key is either missing, incorrectly formatted, or invalid, preventing successful authentication with the Anthropic API.
  fix Set the `ANTHROPIC_API_KEY` environment variable with your valid Anthropic API key (e.g., `os.environ["ANTHROPIC_API_KEY"] = "sk-ant-..."`) or pass it directly when initializing `ChatAnthropic` (e.g., `llm = ChatAnthropic(api_key="sk-ant-...")`).

• ModuleNotFoundError: No module named 'langchain.llms' or ImportError: cannot import name 'AnthropicLLM'

  cause The import path for Anthropic models has changed with the modularization of LangChain, or the `langchain-anthropic` package is not installed.
  fix Ensure you have installed the `langchain-anthropic` package (`pip install -U langchain-anthropic`) and use the correct import path: `from langchain_anthropic import ChatAnthropic`.

• anthropic.RateLimitError: Number of concurrent connections has exceeded your rate limit.

  cause Too many requests are being sent to the Anthropic API within a short period, exceeding the rate limits imposed by Anthropic for your account.
  fix Implement retry logic with exponential backoff using `tenacity`, or leverage LangChain's built-in rate limiting features. Consider requesting a rate limit increase from Anthropic if necessary.

• anthropic.BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages: Unexpected role "ai". Allowed roles are "user" or "assistant"'}}

  cause The message list sent to the Anthropic API includes a message with an unsupported role (e.g., 'ai' when Anthropic expects 'assistant') or an incorrectly ordered system message.
  fix Ensure that `SystemMessage` is the first message in the list, and other roles are correctly mapped to 'user' (`HumanMessage`) or 'assistant' (`AIMessage`) as expected by the Anthropic API, or construct messages carefully following Anthropic's message API format.

• anthropic.BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'This model does not support assistant message prefill. The conversation must end with a user message.'}}

  cause Certain newer Anthropic models (e.g., Claude Opus) do not support starting a new turn with an `AIMessage` (assistant message prefill). If the message history passed to `ChatAnthropic` ends with an AI message, it will cause this error.
  fix Ensure that the final message in the input list to `ChatAnthropic.invoke()` is always a `HumanMessage`, or explicitly manage the conversation history to avoid ending with an AI message for models that don't support prefill.

Warnings

• deprecated The `AnthropicLLM` class for text completion models is considered legacy. For modern Anthropic models like Claude 3, it's strongly recommended to use `ChatAnthropic` instead, which supports the chat-based API.
  Fix: Migrate from `AnthropicLLM` to `ChatAnthropic` and adapt your code to use message-based input/output rather than string-in/string-out.
• breaking The default value for the `max_tokens` parameter in `langchain-anthropic` changed with LangChain v1. Previously, it defaulted to 1024. Now, it defaults to higher values based on the specific model's `max_output_tokens` profile. If your application relied on the old default, it might now consume more tokens or exhibit different truncation behavior.
  Fix: Explicitly set `max_tokens=1024` (or your desired value) when instantiating `ChatAnthropic` or `AnthropicLLM` if you need to maintain the old behavior.
• gotcha It is crucial to set the `ANTHROPIC_API_KEY` environment variable for authentication. Without it, your application will fail when trying to connect to Anthropic models.
  Fix: Ensure `os.environ["ANTHROPIC_API_KEY"]` is set to your valid Anthropic API key before initializing any Anthropic model classes. Consider using a `.env` file and `python-dotenv` for local development.
• breaking LangChain v1 introduced significant architectural changes, including a simplified `langchain` package namespace, revised import paths, and new patterns for agents and tools. Code written for `langchain==0.x` might require substantial modifications.
  Fix: Refer to the official LangChain v1 migration guides and update your imports and component usage according to the new patterns, focusing on partner packages like `langchain-anthropic` for model integrations.
• gotcha When using built-in Anthropic tools like the 'web fetch' tool, be aware of security implications. Enabling web fetching in environments where Claude processes untrusted input alongside sensitive data can pose data exfiltration risks.
  Fix: Only use tools like 'web fetch' in trusted environments or when handling non-sensitive data. Implement strict input validation and access controls.

Install

• pip install langchain-anthropic Install package

Imports

• ChatAnthropic

  from langchain_anthropic import ChatAnthropic

  This is the recommended class for interacting with Anthropic's chat models (e.g., Claude 3, Claude 2.1).
• AnthropicLLM
  wrong:

  from langchain.llms import Anthropic

  correct:

  from langchain_anthropic import AnthropicLLM

  AnthropicLLM is a legacy class for text completion models (e.g., older Claude 2 models) and is generally deprecated in favor of ChatAnthropic. Older LangChain versions used 'from langchain.llms import Anthropic'.

Quickstart

This quickstart demonstrates how to instantiate and use the `ChatAnthropic` model to get a response from a Claude chat model. It highlights the importance of setting the `ANTHROPIC_API_KEY` environment variable.

import os
from langchain_anthropic import ChatAnthropic
from langchain_core.messages import HumanMessage, SystemMessage

# Set your Anthropic API key as an environment variable
# os.environ["ANTHROPIC_API_KEY"] = "YOUR_ANTHROPIC_API_KEY"

# Ensure the API key is set
if not os.environ.get("ANTHROPIC_API_KEY"):
    raise ValueError("ANTHROPIC_API_KEY environment variable not set.")

model = ChatAnthropic(model="claude-3-sonnet-20240229", temperature=0)

messages = [
    SystemMessage(content="You are a helpful AI assistant."),
    HumanMessage(content="What is the capital of France?"),
]

response = model.invoke(messages)
print(response.content)