Making RAG accurate

The pipeline is fully built and it still gets things wrong: outdated policies, vague questions, rules buried in long context. Employees are starting to notice.

This chapter: When a RAG answer is wrong, how do you tell whether retrieval or the model failed, and which lever fixes it?

You will learn

Measure first: retrieval recall vs. answer quality
Seven levers: top-k, metadata filters, query rewriting, reranking, hybrid search, overlap and chunking strategy
Why getting the right text is not the same as using it well

Words you’ll meet (5)

Recall@k: Did the passage with the answer make it into the top k results? Measures retrieval, not the answer.
Top-k: How many chunks the retriever hands over. Too few misses facts; too many buries them.
Metadata filter: A rule on document properties (date, status, team) applied before ranking.
Reranker: A second, more careful model that re-scores the retrieved chunks and reorders them.
Hybrid search: Blending meaning-based search with exact keyword matching.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 17

Your pipeline from the last chapter is live. A consultant booking a trip asks the assistant about the hotel limit.