Query Rewriting

Where we left off

Lesson 5's filtering fixed a case where the right chunk was retrieved but outranked. That trick has a hard limit: if every retrieved chunk grades not-relevant, filtering leaves you with an empty context, no better than the naive answer. This lesson adds the other half of Beginner-tier correction: when grading comes back all-negative, rephrase the query itself and try again.

A named simplification, up front

This is the internal rewrite loop Lesson 1 already named as a simplification of the paper. Yan et al. 2024's actual response to a low-confidence grade is external web search, a different knowledge source entirely. Rewriting the query and searching the same internal corpus again is the easier version to build first, and it's genuinely useful for one specific failure mode (bad wording), but it cannot do what real external search does (find an answer that was never in the corpus). Lesson 22 builds that real branch.

The code, piece by piece

QUESTION = "What is the capital of France?"

Deliberately not a wording problem. None of the five fixture notes have anything to do with France, this question is unanswerable from this corpus no matter how it's phrased. That's the point: this lesson's job is to show the rewrite mechanism honestly, including what it can't do.

REWRITE_PROMPT = """The question below was just asked against a small \
personal notes collection (topics: a home weather station, a garden, a \
pizza dough recipe, a bookshelf, and cello practice), and none of the \
retrieved passages were relevant. Rewrite the question...

Telling the rewriter what topics the corpus actually covers (instead of rewriting blind) gives it a real shot at fixing a wording mismatch, the same idea as a human search user narrowing a query after a bad first result. It's also, implicitly, information Gemini can use to notice a question is off-topic entirely rather than off-topic due to wording.