Retrieval finds candidate information relevant to a need. The returned items still require provenance, access checks, interpretation, and sometimes contradiction handling.

Retrieval Selects Evidence, Not Truth

Chunking divides sources into retrievable units. Good boundaries preserve claims, definitions, qualifiers, and provenance while avoiding fragments too broad or too small to answer reliably.

Chunking Preserves Answerable Context

An embedding maps content to a numeric vector so a distance measure can approximate similarity. Nearby vectors may still differ in truth, authority, time, or required access.

Embeddings Encode Similarity, Not Meaning Guarantees

In this situation—finding policy passages to answer an employee question—which choice best applies “Retrieval Selects Evidence, Not Truth”?

Ranking orders candidates using signals such as lexical match, vector similarity, authority, freshness, graph distance, and access. The weights should reflect the task rather than one universal score.

Learn Before

Related