Project: RAG Reranker

Embedding retrieval gives you passages that are close to the query in vector space. That is not the same as passages that answer it. Ask about a refund policy and you will reliably get back the shipping policy, the cancellation policy, and the terms of service, because they all live in the same neighbourhood.

So put a relevance pass in between. Retrieve twenty by embedding, ask one Noul per passage whether it actually contains an answer, keep what survives, and hand only those to the generating model.

retrieve top 20 by embedding
    ↓
one noul per passage: "does this answer the question"
    ↓
keep passages above threshold  (typically 4-8 survive)
    ↓
generate with the survivors as context

This is the direct countermeasure to context rot, which the Jagged Edges course covers in its own lesson. Sixteen irrelevant passages in a context window are not neutral padding. They are distractors that pull the generation around, and they make a bad answer nearly impossible to debug because you cannot tell which passage caused it.