Tools

it's a prompt limitation, only so much text can fit in the input

Alternatively, you can use an LLM with a larger input size (gpt-3.5-turbo-16k, glt-4-32k), but performance still degrades as the tool selection may be less accurate the more their is

The only other option I can think of is setting up hierarchical agents. Since agents inherit the index base class, they can be used as query engine tools actually

Oh I see

So I would make an agent that uses a bunch of query engine tools as its own query engine tool, interesting

Will explore that perhaps

Yea, basically agents calling agents, it's a neat concept

Yeah the only thing we'd have some difficulty with is organizing structure, since its easier to track which tool is being used to identify which product is being asked about

But if we abstract it to agents, then we'd have to go one step further in that abstraction to identify which agent (we'd have to find a way to name these agents that club some tools), then identify which tool the selected agent is using

@Logan M What does a QueryPlanning tool really do?

How is it different from the FnRetrieverOpenAIAgent that uses an ObjectIndex over tools?

tbh I have no idea what it does or how it works lol

That's one part of the code ive never taken the time to understand. At a high level it creates a plan which is a sequence of queries. It feels pretty beta.