Choose chunks for the material
Use structural boundaries where possible. Very small chunks can lose context; very large chunks can dilute the relevant detail. Preserve table headers and headings. Tune overlap using evaluation rather than assuming one chunk size works everywhere.
Combine complementary retrieval methods
Lexical retrieval can find exact product codes and names. Dense retrieval can match related meanings without identical wording. Hybrid retrieval combines signals; reranking then reorders candidates. A reranker cannot recover a document that was never retrieved.
Diagnose the stage that failed
Measure whether the correct evidence entered the candidate set, survived reranking, and reached generation. Then evaluate whether the answer used it correctly. More retrieved text can add distractions and cost rather than improving the result.
A worked example
Check parsing and indexing, include an exact-match lexical signal, then evaluate candidate recall and reranking. Confirm that the relevant table row and its header reach the answer context.
The point: Fix the failing stage before changing the prompt.
Before you take the exam
Explain these concepts in your own words, work through a different example, and describe why the tempting alternatives fail. Use the references to go deeper.