Source: Notebook experiment 23: Corrective RAG
Runnable scaffold that grades retrieved chunks and falls back to web search when they are weak.
Contents
- What: runnable experiment in the owner’s public Notebook repo
- Author: Loic (coolloic)
- Date: undated (read 2026-11-14)
- Original: experiments/23-corrective-rag
Summary
Retrieves from FAISS, asks the model to grade relevance, and falls back to a stub web search when the grade fails, before answering.
Key claims
- Corrective RAG grades retrieval and takes a corrective action (broaden the query, search the web) when it is weak.
- It targets open-domain questions where the local index may not hold the answer.