LlamaIndex
Log in

Find answers from the community

Updated last year

What would be the best way for me to

At a glance
nno_dice
·
What would be the best way for me to build an agent that avoids the following pitfalls:
  • Selects the wrong tools (when there are several to be confused by)
  • Selects the correct tool initially, but doesn't realize or know to use the info from the first tool as context for another tool.
While being capable of:
  • Route/response planning
  • Leverage multiple tools in its routes, leveraging the response from one tool as the context for another.
We're likely going to have more than 10+ tools soon, so I feel these problems are likely if not addressed.
S
n
8 comments
SSeldo
Which LLM are you using?
nno_dice
ChatGPT 3.5-turbo or GPT 4
SSeldo
You'll probably have noticed already that GPT-4 is a lot better at tool use than 3.5
SSeldo
I wish I had a better answer than "the better the LLM, the more capable it is at tool use" but that's basically my experience.
SSeldo
Though you can get a certain degree further with good tool descriptions and prompt engineering.
nno_dice
Thats a good answer actually.

Our Agent itself uses 3.5 but in other implementations we've used 4. Probably a novice mistake not to compare the two (we lean on 3.5 more than 4 for cost reasons). But if its better at tool selection then thats a justifiable cost. I'll pivot to that.
SSeldo
👍
nno_dice
My current thoughts are also to:
  • Use a reactive agent
  • If possible use an agent that auto-retrieves the tools when a prompt is given, rather than having static tools at creation.
I read about bullet point 2 but haven't dug deep. Not sure if the 1st and 2nd agent types are compatible though.
Add a reply
Sign up and join the conversation on Discord