add reflexion agent

[jerryjliu]

No description provided.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[nerdai]

Was wondering if it made sense to add a reflection auxiliary method to be used within run_step? That way we can handle orthogonal reflection and run_step logics simultaneously. Specifically, if we do that then "CRITIC" can be a reflection implementation that can be used with any run_step logic? Not sure if that makes sense, but just throwing it out there as a thing to perhaps consider. CC: @logan-markewich

[nerdai]

Was wondering if it made sense to add a reflection auxiliary method to be used within run_step? That way we can handle orthogonal reflection and run_step logics simultaneously. Specifically, if we do that then "CRITIC" can be a reflection implementation that can be used with any run_step logic? Not sure if that makes sense, but just throwing it out there as a thing to perhaps consider. CC: @logan-markewich

For CRITIC, I probably will just implement it for now within run_step to move this along. Self-Reflection can be viewed as a a special case to CRITIC (i.e., no tool use in critique phase, and just rely on the LLM itself to perform the reflection).

[logan-markewich]

Yea I was trying to think of a generic way to do this. Adding a "reflection" method to the agent feels a little too specific though 🤔 unless you meant just a general utils function somewhere (then Maybe that's nice, like a little tool box of agent reasoning tools when building a custom loop?)

[nerdai]

I agree having a public method reflection feels a bit too much. Was thinking of it to be a private method, but maybe thats not much different.

I like the sounds of the general utils approach for agents to use. How would you envision implementing something like this?

[nerdai]

Line #60.        return reflection, ChatMessage.from_str(feedback_str, role="user")

Should the reflection take on the role of "user" or "assistant"?

Reply via ReviewNB

[logan-markewich]

I think its supposed to mimic feedback from a human/3rd-party, so user makes sense to me 🤔

[nerdai]

Yeah, just about 10-15 mins ago I had to similar things for CRITIC implementation lol. This makes more sense to me now -- my b.

[logan-markewich]

@nerdai Yea I don't know how feasible or extensible it is, but having a set of tools that have the same API interface to use when designing run_step loops might be useful

I tried doing this at one point, but everything is use-case specific that I wasn't finding it easy to generalize. Maybe a common API interface is just not possible

[nerdai]

I've adapted the self reflection class and ported over to #13108 instead. I've listed you (@jerryjliu) as the author for it in the llamahub metadata 😀

As such, I'm closing this PR!

[jerryjliu]

I've adapted the self reflection class and ported over to #13108 instead. I've listed you (@jerryjliu) as the author for it in the llamahub metadata 😀

As such, I'm closing this PR!

lol don't worry about attribution i'm going to say you made it on the outgoing socials

[nerdai]

I've adapted the self reflection class and ported over to #13108 instead. I've listed you (@jerryjliu) as the author for it in the llamahub metadata 😀
As such, I'm closing this PR!

lol don't worry about attribution i'm going to say you made it on the outgoing socials

lol i figured as much -- excuse my pedantic nature haha