Why it matters
RAG architecture choices lock in latency, cost, isolation, and debuggability for years. Monolith versus microservices, shared versus per-tenant indexes, and synchronous versus async ingestion all affect governance enforcement points.
How it works
Separate indexing from query services so content updates do not require model redeploys. Use collection-scoped indexes or strong metadata filters for isolation. Place policy gates before and after retrieval.
Design citation services that map chunk IDs to stable URLs and version hashes.
Example
Nintendo splits indexing workers from the chatbot API. Help Center updates enqueue re-index jobs per collection. Query nodes read only committed index snapshots so customers never see half-updated refund rules mid-deploy.
Common mistakes
- 1Coupling embedding model changes with immediate production cutover and no regression suite
- 2Single shared reranker trained on internal docs applied to customer queries without filters
- 3No health checks on index freshness lag
Definitions are useful. Governed knowledge is better.
Wiki helps you turn these concepts into a real system for the knowledge behind your AI.
See Wiki in action