Choose a retrieval repair from the first place the expected evidence disappears. Missing text points to ingestion, severed meaning points to chunking, excluded documents point to metadata or filters, and buried relevant passages point to ranking or query behavior.
Use the failed passage to identify the broken control
Inspect the indexed representation of the expected source. If the decisive sentence is absent, repair extraction or ingestion. If it exists but depends on a heading, table label, or preceding definition that was split away, test a boundary-aware chunk. If the chunk is coherent but never eligible for retrieval, inspect tenant, date, product, permission, and document-type metadata before touching semantic search.
When the correct passage is eligible but ranks below broad or duplicated material, compare lexical, vector, and hybrid retrieval on the same query. Query rewriting can help vocabulary mismatch, while deduplication and parent-child retrieval can reduce repeated fragments. Each change should be tested against the same case set because larger chunks or wider recall may improve one query while flooding another with irrelevant context.
Keep repair evidence beside the configuration
Record the exact index version, embedding model, distance metric, filters, top-k, reranker, and query transforms for every run. Save returned document identifiers and passage text, not only aggregate scores. A result table should let another engineer reproduce why a case passed and see whether the source rose because of the intended change or because unrelated configuration drift occurred.
Stop when the bounded acceptance cases pass without an unacceptable regression. Do not infer corpus-wide correctness from a small set. Unresolved cases belong in a visible ledger with the source gap, attempted changes, and next diagnostic. That is more useful than a final report that labels every non-passing case as an unspecified model problem.
Where the service stops
Reality Contact, LLC evaluates and repairs the bounded retrieval path, but does not certify model accuracy, source truth, security, compliance, or correctness for queries outside the agreed cases. The buyer approves the source authority and acceptance cases, reviews the reported limitations, and decides whether and how to deploy the changed retrieval configuration. This is software evaluation and implementation; it does not replace legal, financial, medical, security, compliance, or professional advice. No result establishes corpus-wide correctness, source truth, or reliable behavior for queries outside the buyer-approved case set.
Sources: Pinecone search relevance and chunking guidance; Weaviate hybrid search documentation.