community: Add tool_calls support for ChatTongyi Model, allowing it to be used as the LLM of the lastest tool calling agent

[liushuaikobe]

• Description: Added tool calls support for the ChatTongyi Model (which was officially supported recently), now the ChatTongyi can be used as the LLM of a tool calling agent seamlessly.

example usage:

llm = ChatTongyi(model_name="qwen-max")
tools = [TavilySearchResults(max_results=1)]
prompt = ChatPromptTemplate.from_messages(
  [
      (
          "system",
          "You are a helpful assistant. Make sure to use the tavily_search_results_json tool for information.",
      ),
      ("placeholder", "{chat_history}"),
      ("human", "{input}"),
      ("placeholder", "{agent_scratchpad}"),
  ]
)
# Construct the Tools agent
agent = create_tool_calling_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)
agent_executor.invoke({"input": "what is LangChain?"})

or simply:

class GetWeather(BaseModel):
  """Get the current weather in a given location"""

  location: str = Field(
      ...,
      description="The city and state, e.g. San Francisco, CA",
  )
llm = ChatTongyi(model_name="qwen-max")
llm_with_tools = llm.bind_tools([GetWeather])
llm_with_tools.invoke("what is the weather like in HangZhou, China")

• Dependencies: no new dependencies required.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		May 7, 2024 8:26am

[ccurme]

Thanks for this @liushuaikobe. This appears duplicated with #20725 so I'm closing it for now. Would appreciate your review and/or testing on that PR if you can.

[liushuaikobe]

@ccurme Thank you for your review and I'm also grateful for #20725!

Actually, this PR employs a more "langchain" approach and aligns better with the original design of langchain..

1. It utilizes tool_call_chunks in AIMessageChunk for handing the streaming output of AI message with tool calls. It also defines an inner subclass of AIMessageChunk to handle the limitations of Tongyi LLM API appropriately.
2. The utility method message_chunk_to_message from the core library is used when converting a message or message chunk into a dictionary.
3. Explicitly passes the incremental_output parameter when calling the bind method for tool binding, clearly demonstrating the requirements of the Tongyi LLM API. Moreover, I have also included additional comments for clearer explanations and better code understanding.

Hence, this PR entails minimal code changes while fully supporting tool calling agents with the Tongyi LLM.

I would greatly appreciate it if this PR could be reopened and reviewed once again.

[ccurme]

@liushuaikobe thanks for the work you did here. We merged #20725. Would appreciate your improvements to that implementation.

[jlcoo]

@liushuaikobe @ccurme

the ChatTongyi can't deal tool response?