Introduce dspy.Refine, dspy.BestOfN and fix thread safety for dspy.ChainOfThoughtWithHint
#1959
Conversation
chenmoneygithub
force-pushed
the
dspy-refine
branch
from
a1a6bb2 to
a8e3f19
Compare
|
I'm really interested in this... I mean in terms of specs! |
There was a problem hiding this comment.
Really cool PR!
A project I am working on used to leverage the dspy.Suggest/Assert functionality (before the recent deprecation), and this looks like an amazing substitute.
I added some comments/feedback/questions, which I hope will help in building this. Let me know if I can be of service :D
|
@zbambergerNLP Thanks for your comments! Usually we don't recommend community dropping comments on PR, a much better way to contribute is trying this API out in your workflow, and let us know what's good/bad (ideally with a colab link for reproducing). We will have this API merged and released in 1-2 weeks, stay tuned! |
chenmoneygithub
requested review from
zbambergerNLP and
okhat
and removed request for
zbambergerNLP
|
Thanks a lot @zbambergerNLP ! @chenmoneygithub , @zbambergerNLP is not just a community member, he's been building some cool things in the broader team :D |
dspy/refine/refine.py
Outdated
| # last_trace. | ||
| predict.forward = MethodType(partial_forward, predict) | ||
|
|
||
| def _undo_patch_predict_call_with_feedback(self, named_predicts): |
okhat
changed the title
dspy.Refinedspy.Refine, dspy.BestOfN and fix thread safety for dspy.ChainOfThoughtWithHint
…into pr/chenmoneygithub/1959
dspy.Refineis a new API we are adding to support feedback-based retry. More details coming soon.