Modify context before LLM

[tpy37]

Thank you for the great package. I am looking for a method to change the context before sending it to the LLM in MultimodalAgent class. I think it exists in VoicePipelineAgent, and I am wondering how I could implement it with OpenAI realtime API.

[tpy37]

I would be happy if you could implement the RAG part of the MultimodalAgent, as in the documentation! :)
before_llm_cb=_enrich_with_rag

[davidzhao]

it's a bit difficult with multimodal agent, because it's from voice input directly to voice output.

the way to handle RAG is with function calling. If you are defining a function for the LLM to look up information with the user's query, it should be straight forward to pick up the function call and return the RAG results that way

[tpy37]

Thank you very much David!
I see... I was trying the function call, but changing to tool "required" seem to completely halt the process and break the conversation, so I stopped using it.

Since we have the transcribed text from the user's audio in the
openai.realtime.RealtimeResponse

I was thinking that we could analyze the transcribed texts, and then use that to do RAG, and send the results async to the openai api as texts?
const event = { type: 'conversation.item.create', item: { type: 'message', role: 'user', content: [ { type: 'input_text', text: 'Hello!' } ] } }; ws.send(JSON.stringify(event)); ws.send(JSON.stringify({type: 'response.create'}));
https://platform.openai.com/docs/guides/realtime?text-generation-quickstart-example=text

Just some thoughts...

[prashantmetadome]

@tpy37 can you guide me on how to change context in VoicePipelineAgent?

[tpy37]

I am sorry, I haven't done it myself in VoicePipeline. But the example is available in https://docs.livekit.io/agents/voice-agent/voice-pipeline/#modify-context-before-llm

I think there was also example to send it to RAG using this before_llm_cb in one of the Github repository:
examples/voice-pipeline-agent/simple-rag/assistant.py

Hope it helps!

[prashantmetadome]

@tpy37 thank you very much, it does help. But I am still unsure.. my use case is I want to manipulate the prompt based on a tool call.

To go deeper in the specific requirement, the conversation has outgrown the current prompt and has gone into different territory that need to be handled by a different prompt.

I do not need to manipulate the prompt inside the tool call but if I can just extract some metadata from tool call and access it in the callback function should solve the problem but I am not sure how can I do that.
@davidzhao any help would be appreciated.