please support langchain AgentExecutor stream function about ai HOT 10 CLOSED

Did anyone get this to work?

For example

const result = await agent.stream({
	input: currentMessageContent,
	outputKey: "output",
	chat_history: memory.chatHistory.getMessages()
});

returns an IterableReadableStream<ChainValues> which isn't an accepted input for LangChainAdapter, even after the #2340

Trying what ivanvanderbyl did, didn't work for me either. Though it seems to be going in the right direction. Is this what we are expected to do? As the documentation hasn't been updated to state that.

For the most part I have been able to make this work with:

const agentStream = await agentExecutor.streamEvents(
  { messages, input: lastMessage.content },
  { version: "v2" },
);

const stream = new ReadableStream({
  async pull(controller) {
    for await (const { event, data } of agentStream) {
      if (event === "on_chat_model_stream") {
        const msg = data?.chunk as AIMessageChunk;

        if ((msg.tool_call_chunks?.length ?? 0) > 0) {
          // console.log("Tool calls", msg.tool_calls);
        } else if (msg.content) {
          controller.enqueue(msg.content as string);
        }
      }
    }
    controller.close();
  },
});

const aiStream = LangChainAdapter.toAIStream(stream);
return new StreamingTextResponse(aiStream);

You need to use streamEvents():

const agentStream = await agentExecutor.streamEvents(
	{ messages, input: lastMessage?.content ?? "" },
	{ version: "v2" },
);

const aiStream = LangChainAdapter.toAIStream(agentStream);
return new StreamingTextResponse(aiStream);

from ai.

👍 🙏

from ai.

+1 would be very helpful

from ai.

👍 🙏

from ai.

For the most part I have been able to make this work with:

const agentStream = await agentExecutor.streamEvents(
  { messages, input: lastMessage.content },
  { version: "v2" },
);

const stream = new ReadableStream({
  async pull(controller) {
    for await (const { event, data } of agentStream) {
      if (event === "on_chat_model_stream") {
        const msg = data?.chunk as AIMessageChunk;

        if ((msg.tool_call_chunks?.length ?? 0) > 0) {
          // console.log("Tool calls", msg.tool_calls);
        } else if (msg.content) {
          controller.enqueue(msg.content as string);
        }
      }
    }
    controller.close();
  },
});

const aiStream = LangChainAdapter.toAIStream(stream);
return new StreamingTextResponse(aiStream);

from ai.

#2340

from ai.

fixed in main

from ai.

Did anyone get this to work?

For example

const result = await agent.stream({
	input: currentMessageContent,
	outputKey: "output",
	chat_history: memory.chatHistory.getMessages()
});

returns an IterableReadableStream<ChainValues> which isn't an accepted input for LangChainAdapter, even after the #2340

Trying what ivanvanderbyl did, didn't work for me either. Though it seems to be going in the right direction. Is this what we are expected to do? As the documentation hasn't been updated to state that.

For the most part I have been able to make this work with:

const agentStream = await agentExecutor.streamEvents(
  { messages, input: lastMessage.content },
  { version: "v2" },
);

const stream = new ReadableStream({
  async pull(controller) {
    for await (const { event, data } of agentStream) {
      if (event === "on_chat_model_stream") {
        const msg = data?.chunk as AIMessageChunk;

        if ((msg.tool_call_chunks?.length ?? 0) > 0) {
          // console.log("Tool calls", msg.tool_calls);
        } else if (msg.content) {
          controller.enqueue(msg.content as string);
        }
      }
    }
    controller.close();
  },
});

const aiStream = LangChainAdapter.toAIStream(stream);
return new StreamingTextResponse(aiStream);

from ai.

@hiddengearz can you point me to the ChainValues source code / docs for the latest langchain? I can take a look add adding the relevant support.

from ai.

Yikes, I updated the @ai-sdk/svelte package but forgot to update ai. I won't get a chance to test it until the weekend but that fixed the build errors.

Thanks for the quick responses and great job!

from ai.