[Deprecated] Experimental Anthropic Tools Wrapper

danger

The Anthropic API officially supports tool-calling so this workaround is no longer needed. Please use ChatAnthropic with langchain-anthropic>=0.1.5.

This notebook shows how to use an experimental wrapper around Anthropic that gives it tool calling and structured output capabilities. It follows Anthropic’s guide here

The wrapper is available from the langchain-anthropic package, and it also requires the optional dependency defusedxml for parsing XML output from the llm.

Note: this is a beta feature that will be replaced by Anthropic’s formal implementation of tool calling, but it is useful for testing and experimentation in the meantime.

%pip install -qU langchain-anthropic defusedxml
from langchain_anthropic.experimental import ChatAnthropicTools

Tool Binding

ChatAnthropicTools exposes a bind_tools method that allows you to pass in Pydantic models or BaseTools to the llm.

from langchain_core.pydantic_v1 import BaseModel


class Person(BaseModel):
    name: str
    age: int


model = ChatAnthropicTools(model="claude-3-opus-20240229").bind_tools(tools=[Person])
model.invoke("I am a 27 year old named Erick")

AIMessage(content='', additional_kwargs={'tool_calls': [{'function': {'name': 'Person', 'arguments': '{"name": "Erick", "age": "27"}'}, 'type': 'function'}]})

Structured Output

ChatAnthropicTools also implements the with_structured_output spec for extracting values. Note: this may not be as stable as with models that explicitly offer tool calling.

chain = ChatAnthropicTools(model="claude-3-opus-20240229").with_structured_output(
    Person
)
chain.invoke("I am a 27 year old named Erick")

Person(name='Erick', age=27)