Include request's tools configuration in OpenAI LLM calls

[ricwo]

Describe the feature or potential improvement

Description

For debugging agentic systems, it would be really helpful to include the tools field of the LLM call, at least for OpenAI. The LLM response already includes the chosen tool (if any).

Example

My outgoing OpenAI request looks like this (through LlamaIndex):

{
  "messages": [
    "user: How are you doing?"
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "query_engine_tool",
        "description": "Useful for running a natural language query\nagainst a knowledge base and get back a natural language response.\n",
        "parameters": {
          "type": "object",
          "properties": {
            "input": {
              "title": "Input",
              "type": "string"
            }
          },
          "required": [
            "input"
          ]
        }
      }
    }
  ],
  "tool_choice": "query_engine_tool"
}

Langfuse only captures the following (Tracing -> Generations):

[
    {
        "role": "user",
        "content": "How are you doing?",
        "additional_kwargs": {}
    }
]

Langfuse correctly captures the assistant's response including the choice of tool:

{
    "tool_calls": [
        {
            "id": "call_VCTCLBgPo6uSEzv3yWWCnIqR",
            "type": "function",
            "function": {
                "name": "query_engine_tool",
                "arguments": {
                    "input": "How are you doing?"
                }
            }
        }
    ]
}

Request

Might there already be a config parameter that enables this? If not, is there any chance this can be included and captured through LlamaIndexCallbackHandler?

Additional information

No response

[marcklingen]

Thanks for sharing, this should be captured, can you open an issue with code to reproduce this and a public link to a trace of this?