Search records

Resolve the requested profile, compile mandatory Corpus/Organization prefilters and any requested filter, obtain candidates, then canonically hydrate and reauthorize every returned segment. Lexical-first records remain eligible without embeddings; semantic-only queries require vector coverage. Ordinary hybrid searches use keywords for Corpora whose served vectors are unavailable, reporting their ids in retrieval_profile.degraded. Other Corpora keep their hybrid ranking. Explicit semantic requests remain 422 unsupported_search when vector search cannot run. Unknown coverage alone does not disable a routed space. Profile selection does not change access/currentness rules. When a retrieval plugin is pinned, it ranks. It asks the engine for candidates in up to three rounds and returns its ranking, which may hold only candidates the engine served in this search, each already authorized and hydrated. Each API process admits at most 64 active searches across callers. At capacity it immediately returns retryable 503 search_unavailable with Retry-After before storage or plugin work. A query encoding model that cannot answer returns retryable 503 model_unavailable, including the legacy inference service and the ingestion plugin that owns the requested vector space.