docs/tutorials/agents/

[giscus[bot]]

docs/tutorials/agents/

This guide assumes familiarity with the following concepts:

https://python.langchain.com/docs/tutorials/agents/

[ganeshghag]

assuming there are multiple agents tools being bound into the model, how does "a qualitative input NLP question" trigger appropriate tool? there could be multiple tools bound which are capable of providing the answer?

[ZcyhFromC]

Hi gang. With multi tools available, agents answer the question in three steps:

1. Let llm leveraging the metadata of tools to reason out which tool to use. Some description will work.
2. Let python code invoke the tool and feed the response back to llm
3. llm final output
   The key is that llm will first act as a router.

[ganeshghag]

Thanks I read in detail about "tool calling" and now realize your reply is significant

[ganeshghag]

Also, how can I rely on the third-party tool implementations, to give my users proper responses.
I cannot be testing for all possible inputs to these tools.
With above ambiguity, I am relegated to writing agents and my own custom tools

Tools appear to be wrappers over remote REST API, with a technical interface like runnable so that the agent can auto-invoke these tools

its a meek attempt to standardize AI agents with public REST APIs

[shan-max]

yes i am there , and my name shanawaz.

[shaojun]

can you show how to use trim_messages to limit the messages in Saver in this agent scenario?

[Bhargav2525]

I am not getting response. It seems llm is not using that tool.
response = agent_executor.invoke(
{"messages": [HumanMessage(content="whats the weather in sf?")]}
)
response["messages"]
[HumanMessage(content='whats the weather in sf?', additional_kwargs={}, response_metadata={}, id='15979f0b-4a79-46df-8d28-43bc289aa6'),
AIMessage(content="I apologize, but I don't have a specific tool available to check the current weather in San Francisco. However, I can use the search function to find recent weather information for you. Would you like me to search for the current weather in San Francisco?", additional_kwargs={}, response_metadata={'ResponseMetadata': {'RequestId': 'c3d31db4-d840-49eb-8767-48eba5f89', 'HTTPStatusCode': 200, 'HTTPHeaders': {'date': 'Fri, 13 Dec 2024 07:26:24 GMT', 'content-type': 'application/json', 'content-length': '439', 'connection': 'keep-alive', 'x-amzn-requestid': 'c3d31db4-d840-49eb-8767-48e9cb9'}, 'RetryAttempts': 0}, 'stopReason': 'end_turn', 'metrics': {'latencyMs': [1670]}}, id='run-d819b59d-2a1e-4d83-85a3-e714a5c2-0', usage_metadata={'input_tokens': 404, 'output_tokens': 55, 'total_tokens': 459})]

[YanzhongSu]

How was your agent_executor initialised?

[yanjiecdh]

hi, Can this method execute multiple tools consecutively, and can the execution results of the tools be passed on to the next tool