What Is Corrective RAG

Where we left off

naive_rag (course 1) built the baseline shape: chunk, embed, retrieve, generate. That course's own Lesson 14 and Lesson 22 (choosing_k_and_thresholds, metadata_filtering_with_chromadb) already showed the specific crack in that shape: top-k retrieval always returns its k best-scoring chunks, even when none of them are actually relevant, and generation has no way to know the difference between "these chunks answer the question" and "these chunks merely scored highest." A similarity threshold helps when nothing scores high enough to be worth using, but it does nothing for the harder case, a wrong chunk that scores high because it's topically close, just not actually the right one. Corrective RAG is a full architecture built to close exactly that gap.

What Corrective RAG adds

Yan et al. 2024, "Corrective Retrieval Augmented Generation" (arXiv:2401.15884), proposes a retrieval evaluator: a lightweight model that grades each retrieved chunk's confidence before generation ever sees it, into three buckets:

• Correct: keep it, but refine it further (the paper decomposes the chunk into strips and recomposes only the relevant ones).
• Ambiguous: keep the refined version, but also treat this as a signal the retrieval might be incomplete.
• Incorrect: discard it, and fall back to an external knowledge source (the paper uses web search) instead of generating from it.

That's the whole idea: retrieval alone can't tell good matches from plausible-looking bad ones, so add a step whose only job is grading what came back, then act differently depending on the grade.

Two simplifications this course makes, and where they're closed

Building the full paper on day one skips the parts that make each piece legible on its own. This course makes two deliberate simplifications early, both named here so they don't surprise you later if you read the paper directly:

1. Beginner's rewrite loop queries the same corpus. Lessons 6-7 handle a "nothing retrieved is relevant" grade by rephrasing the query and re-retrieving from the same internal document set. The paper's actual response to a low-confidence grade is external web search, a fundamentally different corpus. Lesson 22 (Advanced tier) builds that real branch, and is explicit that it, not the Beginner loop, is what "Corrective RAG" means in the literature.
2. Beginner grades chunks as binary relevant / not-relevant. Lesson 3 asks Gemini for a single yes/no call per chunk. The paper's full three-way correct / ambiguous / incorrect confidence grade needs strip-level refinement to have anything useful to do with "ambiguous," which doesn't exist until Lessons 10-11. Lesson 12 upgrades the binary grade to the real three-way bucket once that machinery is in place.

The code, piece by piece

This lesson's lesson.py is deliberately small, it asks Gemini the question this whole course is going to keep asking, with no context at all, the same "prove the gap exists before building the fix" move Lesson 1 of naive_rag made. QUESTION is chosen so that naive top-1 retrieval (proven with real embeddings in Lesson 2) confidently returns the wrong fixture file, not just a low-scoring miss, that specific failure shape is this course's entire premise.