Database

corpus stores documents in a pgvector-backed table inside CORPUS_DB_SCHEMA. The pgvector document store creates and manages that table, so the DB role must own the schema.

A small sync_state table (source, cursor, updated_at) tracks per-source progress. The cursor is opaque — each fetcher encodes whatever it needs to resume incrementally (see Sources). Identity is (source, source_uid), so re-ingesting an item overwrites its row rather than adding a duplicate.

The derived enrichment and secret-audit records live in a separate enrichments table, created lazily on first write. The architecture diagram shows how the derived branches relate to the source-of-truth documents.