Optimizing multiple index options for consistent respon...

At a glance

The community members are discussing the best way to combine multiple index options. The original poster tried using the CondensePlusContextChatEngine, but it kept rewriting the input and returning an empty response. The comments suggest using an agent or the query fusion retriever as alternatives. Some community members discuss using an agent with a chat engine, and note that an agent is a type of chat engine that can be given tools to access the index. They also mention the FunctionCallingAgent as an option for LLMs that support tools. There is no explicitly marked answer, but the community members provide several suggestions for combining multiple index options.

ZZachHandley

What's the best way to combine multiple index options? I tried the CondensePlusContextChatEngine and it keeps rewriting e.g. "Hello" into a bunch of different things and then returning an empty response, but most examples in the docs want me to use an index -> chat engine

19 comments

LLogan M

CondensePlusContext will rewrite yes, but it should respond as normal (its only rewriting and then retrieving, and then putting retrieved context + chat history to the llm). You may need to adjust the system prompt

LLogan M

Beyond that, combining multiple indexes, I usually use either an agent or the query fusion retriever

ZZachHandley

can I use the agent with a chat engine?

LLogan M

uuhhh an agent is a chat engine -- you'd give the agent tools that access your index

LLogan M

You could mix them I guess

LLogan M

But then chat history/state will get very confusing

ZZachHandley

Oh, I can just give the agent a chat history? I thought they'd be diff

ZZachHandley

ah I see

ZZachHandley

I basically want a chat engine with the ability to use query tools

LLogan M

yea thatd be an agent then 👍

ZZachHandley

ReaCt would be nice, but it needs the chat history, maybe

ZZachHandley

Plain Text

fusion_retriever = QueryFusionRetriever(
            retrievers=retrievers,
            similarity_top_k=3,
            num_queries=3,
            mode=FUSION_MODES.RECIPROCAL_RANK,
            verbose=True,
        )

so like this

ZZachHandley

with the CondensePlusContext

LLogan M

If you are using an LLM that supports tools (openai, anthropic) I would use the FunctionCallingAgent instead of react

LLogan M

and you could give it that retriever as a tool (or use a query engine as a tool, up to you)

ZZachHandley

what's weird or annoying

ZZachHandley

is that I gave it a pretty basic prompt of who it is and what not

ZZachHandley

can I DM you?

ZZachHandley

related to project

Add a reply

Find answers from the community

Optimizing multiple index options for consistent responses