Hover a node · click to focus · ←/→ steps
Ingest → embed → index → retrieve → generate. Offline + online eval gates block regressions before promote.
When you need this
- Notebook FAISS index became the prod retrieval path
- Need EU-hosted vectors with backup/restore
- Multi-tenant RAG with isolation requirements
- Choosing between dedicated vector DB and Postgres pgvector
Prerequisites
- Document ingestion pipeline (e.g. Unstructured)
- Embedding model + version pinned
- Latency and recall SLOs defined
Tools
Prefer when collections grow past ~1M vectors or filtered search is first-class.
Default for mid-market under ~500k–1M vectors; plan migration path before HNSW RAM blows up.
Assess vs Qdrant/pgvector for EU data residency, exit cost, and filter latency needs.
Trial on real customer docs early - naive text splitters fail on tables and scans.
Steps
- 1
Pick the store for your constraints
pgvector when you already run Postgres and volume is moderate; Qdrant when you need dedicated ANN performance / filtering; Pinecone when managed ops outweighs lock-in. Document the decision.
- 2
Design collections and metadata
Payload filters for tenant, product, language, ACL. Never retrieve across tenants. Store source URI + chunk hash for citations.
- 3
Ingest with versioned embeddings
Pipeline: parse → chunk → embed → upsert. Tag index with embedding model version; rebuild on model change, do not mix vectors.
- 4
Operate: backup, scale, observe
Snapshot/backup schedule, capacity alerts, p95 query latency dashboard. Test restore quarterly.
- 5
Gate quality on the index
RAG eval suite (Langfuse + Ragas) runs against the staging index before every schema or embedding change.
Adoption pitfalls
- Mixing embedding models in one collection
- No ACL in metadata → cross-tenant leaks
- Rebuilding prod index without offline eval
Adoption checklist
- Embedding model version pinned on the index
- Tenant isolation tested
- Backup/restore drill completed
- RAG eval gate on index changes
- Latency SLO dashboard live
SEER REAL assessment / sprint
Assessment checks whether retrieval is a real production system. Sprint typically productionizes one vector store with backups + eval gate for a single RAG product.