Dedicated AI data plane and vector lifecycle

Separation

The AI/vector database is distinct from the transactional runtime database. This keeps semantic indexing, model migrations, rebuilds, and retrieval telemetry from becoming ad hoc state inside the operational system.

Current vector pattern

The August 2026 architecture includes multi-resolution vector storage and secured retrieval paths. The documented current path uses a smaller embedding for candidate generation and a larger embedding for precision ordering, with an alternate high-resolution route also represented.

Provenance

AI chunks carry provenance such as SaaSAppId, AppAccountId, SourceRecordId, SourceModuleId, LocationId, security scope, model/version, chunking version, and hashes/checksums.

Lifecycle

Ingest source → chunk with versioned rules → embed → persist provenance → retrieve inside authorization scope → rerank/precision pass → log result/latency/model path → rebuild when model or source changes.