How to use critic-model in the RAG workflow?

[YuanNing-NEU]

Hello author, I have downloaded the weight of the 3B critic model, and may I ask how to use it for multi-iteration RAG in the code you provided? And how to start the first iteration?

[onestardao]

To run the first iteration with the 3B critic model in a multi-iteration RAG, you’d usually:
Load the critic model alongside your retriever and generator.
After the generator produces an answer, pass it to the critic for scoring or refinement.
Use that scored output to guide the next retrieval/generation cycle.

That said, multi-iteration loops with a critic often run into ProblemMap No.3 — Long Reasoning Chains, where outputs drift across steps. I keep a MIT-licensed reference of these common patterns and tested fixes — if you’d like to avoid those pitfalls, I can share the link.

[YuanNing-NEU]

@onestardao Thank you very much. So you mean I need to first run “Standard” RAG, then pass its results to the critic model？

[onestardao]

Yes, exactly — run the standard RAG step first, then pass results to the critic model for scoring/refinement.

If you want to avoid the common drift issues in long reasoning loops, here’s the MIT-licensed reference I mentioned:
WFGY ProblemMap — Long Reasoning Chains (No.3)

It includes tested fixes and patterns you can adapt directly.