Refinery

Parsed, deduplicated, schema-normalized.

Crawled and API-sourced content is resolved against a canonical schema — entities reconciled, duplicates collapsed, every field validated and stamped with its provenance — so retrieval downstream never has to infer format or origin.

Storage

Append-only. Encrypted. Tenant-isolated.

Refined records are written once and versioned rather than overwritten, held under zero-knowledge encryption with edge replication, and indexed for both vector and keyword retrieval. Any answer can be replayed against the exact record state that produced it.

Terminal

Retrieval-grounded. Citation-bound.

Queries resolve against the indexed store, not model memory: retrieved passages are ranked, bound to their source record IDs, and returned over REST or SQL — so every response traces back to the record it came from.