Support agents with Mistral api

[gingters]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Just with openai agents, I want to be able to build an agent against Mistral-large, which is currently only available through the Mistral API. Currently this is not possible.

Motivation

Mistral-large seems to be capable of producing great reasoning just along the lines of GPT-4. If you want to implement an agent with function calling against mistral-large, you currently are completely blocked by the lack of an agent implementation for the Mistral API.

Proposal (If applicable)

No response

[pierreveron]

The issue I found is that the AIMessage containing the function call that is sent back by Mistral API doesn't have content. As so, in the _stream() and _astream() implementations, the loop on chunks continue and the chunk cannot be converted to a message chunk. My quick fix was to inherit ChatMistralAI and override the implementations. Let me know if that helps! @gingters

class FixedChatMistralAI(ChatMistralAI):

    def _stream(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Optional[CallbackManagerForLLMRun] = None,
        **kwargs: Any,
    ) -> Iterator[ChatGenerationChunk]:
        message_dicts, params = self._create_message_dicts(messages, stop)
        params = {**params, **kwargs, "stream": True}

        default_chunk_class = AIMessageChunk
        for chunk in self.completion_with_retry(
            messages=message_dicts, run_manager=run_manager, **params
        ):
            if len(chunk.choices) == 0:
                continue
            delta = chunk.choices[0].delta
            if not delta.content:
                # This is what changes
                if hasattr(delta, "tool_calls") and getattr(delta, "tool_calls"):
                    # Set the content to empty string to prevent error in _convert_delta_to_message_chunk()
                    delta.content = ""
                else:
                    continue
            chunk = _convert_delta_to_message_chunk(delta, default_chunk_class)
            default_chunk_class = chunk.__class__
            if run_manager:
                run_manager.on_llm_new_token(token=chunk.content, chunk=chunk)
            yield ChatGenerationChunk(message=chunk)

    async def _astream(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
        **kwargs: Any,
    ) -> AsyncIterator[ChatGenerationChunk]:
        message_dicts, params = self._create_message_dicts(messages, stop)
        params = {**params, **kwargs, "stream": True}

        default_chunk_class = AIMessageChunk
        async for chunk in await acompletion_with_retry(
            self, messages=message_dicts, run_manager=run_manager, **params
        ):
            if len(chunk.choices) == 0:
                continue
            delta = chunk.choices[0].delta
            if not delta.content:
                # This is what changes
                if hasattr(delta, "tool_calls") and getattr(delta, "tool_calls"):
                    # Set the content to empty string to prevent error in _convert_delta_to_message_chunk()
                    delta.content = ""
                else:
                    continue
            chunk = _convert_delta_to_message_chunk(delta, default_chunk_class)
            default_chunk_class = chunk.__class__
            if run_manager:
                await run_manager.on_llm_new_token(token=chunk.content, chunk=chunk)
            yield ChatGenerationChunk(message=chunk)