Optimize System/tool directed queries with response from custom RAGs rather than underlying models

[nivemaham]

Describe the bug
When prompted with generic system based questions like "Tell me about yourself" or "What can you do" the system responses with underlying model. This can be misleading to less technical users who think this is not ChatGPT.

To Reproduce
Prompt with "Tell me about your self"

Stack trace

Expected behavior
BioChatter should respond with a concise summary on itself and what can do.

Desktop (please complete the following information):

• Current online demo

Additional context
UX improvements.

[slobentanzer]

Hi @nivemaham,
thanks for raising this, sorry for the delay. Missed it because I was on holiday and then it went under.

You make a valid point, but I am also not sure whether we should instruct the LLM to do so specifically, as it uses up tokens and may bias the model in other ways. We surely would need to benchmark it first. But it is true that UX is going to be mainly users asking "what are you" in the context of LLM interactions. We have the "disadvantage" of being able to operate on any LLM, which means this kind of question would have to be anticipated and encoded for in the system prompt. Another, even more involved approach would be to check for this question (or any kind of usage / UX question) and redirect in an agentic setting to a sysinfo agent. The answer that ChatGPT gave above is not wrong, it is just not the most specific.

We'll have to give this some thought, but not sure whether we should do it for the BioChatter web demo.

[nivemaham]

Hi @slobentanzer Thanks for you response and acknowledgement.
True. I gave it some thoughts too. After navigating few AI/LLM depended tools, I do see that the response can/should be model dependent. I think the ability to support any model is a strength. Indeed, totally agree with the approach. I think this can be a nice use-case to try implement an agentic-interceptor and response with a custom augmented response. (in all honestly, I understand this may not be a high priority).