AI Executive Office — CXO Assistant Platform · View 14 of 30 · 4 · Data
Decisions
- Source permissions are captured at ingestion and carried as security fields on every chunk. They are never re-derived at query time, because the derivation is the part that goes wrong
- The security filter is applied inside the search query, before scoring. Filtering after retrieval leaks nothing but silently returns a worse top-k, which is the failure nobody notices
- Hybrid retrieval — vector plus keyword, fused — because executive questions mix concepts with exact identifiers such as a contract number or a project code that a vector search alone handles poorly
Numbers
- Retrieve 50, rerank to 8 passages. Chunking is per content type, not one global size
- Documents are superseded rather than deleted, so a four-month-old citation still resolves to the version that was cited (view 30)
- Crawl cadence 4 hours; a newly restricted document loses visibility on the next crawl, which is a stated exposure window
Risks
- Permission drift between crawls is a real exposure. Mitigation is a shorter crawl for high-sensitivity libraries and a runtime re-check for anything labelled restricted, at a latency cost
