Why it matters
Retrieval is the lookup step. It decides which evidence the model even sees. Perfect generation cannot recover from missing or wrong chunks.
How it works
Embed or tokenize the query, search indexes, apply metadata filters, rerank, deduplicate, and truncate to context limits. Retrieval policies run before and after ranking.
Example
Nintendo retrieval for refund questions searches only the public-safe Help Center and Refund Policy collections, applies verified-status filter, returns top five chunks, then reranks by section authority metadata.
Common mistakes
- 1Top-k too small for multi-hop questions
- 2No deduplication of overlapping chunks wasting context
- 3Retrieval tuned on internal jargon while customers use plain language
Definitions are useful. Governed knowledge is better.
Wiki helps you turn these concepts into a real system for the knowledge behind your AI.
See Wiki in action