Methodology

The agent first inspects low-cost document metadata and structure. It then searches for candidate sections and reads only the full passages needed for evidence.

Search snippets are discovery aids, not evidence. A full authoritative section must be read before it can support the answer.

The server creates a stable opaque tag for every passage read. The model only copies that tag. Persisted metadata reconnects the citation after a reload.

Active release: health-kb-2026-07-31-31cb166817d6, containing 434 documents, 8,079 sections, and 12,773 pages. New releases are staged and reconciled before activation.