Vector Databases (Deep Workflow)

Vector search is approximate nearest neighbor (ANN) at scale—not magic semantic understanding. Success requires embedding model alignment, index parameters, metadata filters, and evaluation against real queries.

When to Offer This Workflow

Trigger conditions:

• Building RAG, similarity search, dedup, recommendations, anomaly clustering
• Comparing managed vector DB vs pgvector vs search engine kNN
• Recall issues, stale vectors, slow queries, or cost explosions

Initial offer:

Use six stages: (1) problem & metrics, (2) embeddings & schema, (3) index & parameters, (4) hybrid & filtering, (5) operations & cost, (6) evaluation & iteration. Confirm scale (vectors, QPS, dimension) and latency SLO.

Stage 1: Problem & Metrics

Goal: Define what “similar” means for the product—not only cosine similarity.

Questions

1. Query types: short keyword vs long paragraph? multilingual?
2. Precision vs recall preference: legal/medical may need high precision
3. Freshness: how often do vectors change? Real-time upserts? 4 Ground truth: any labeled relevant pairs for eval?

Metrics

• Recall@k, MRR, nDCG when judgments exist; otherwise human spot checks + proxy tasks

Exit condition: Success metric and minimum acceptable recall/latency stated.

Stage 2: Embeddings & Schema

Goal: Stable embedding pipeline with versioning and metadata design.

Embeddings

• Model choice: domain fit (code vs general text); dimension; distance metric (cosine vs dot vs L2)—match DB defaults
• Chunking strategy upstream—bad chunks → bad retrieval regardless of DB

Schema

• Payload/metadata per vector: doc_id, tenant_id, acl, source, timestamps
• Multi-vector per doc (passages) vs single centroid—trade-offs

Versioning

• Re-embed all on model change—plan downtime or dual-write period

Exit condition: ID strategy + metadata filter needs documented.

Stage 3: Index & Parameters

Goal: Pick index type and build params for data size and recall.

Common families (vendor-specific names)

• HNSW: strong latency/recall; memory hungry; tunable efConstruction, M
• IVF: better memory; needs training nlist; probe tuning
• PQ/OPQ: compression—recall hit; good for huge scale

Tuning loop

• Start defaults; sweep parameters with benchmark queries
• Watch insert throughput during index build on large backfills

Exit condition: Benchmark results: p95 latency vs recall at fixed k.

Stage 4: Hybrid Search & Filtering

Goal: Combine vector similarity with structured constraints—most production needs this.

Patterns

• Pre-filter metadata (tenant, date) before ANN when supported—verify filter selectivity
• Hybrid: BM25 + vector with weighted fusion or rerank stage
• Reranking: cross-encoder on top-k candidates—quality boost, latency cost

Pitfalls

• Filtering that leaves too few candidates—empty results despite “similar” existing in other tenants

Exit condition: Query plan documented: ANN → filter → rerank (as applicable).

Stage 5: Operations & Cost

Goal: Reliable ingestion, monitoring, and predictable bills.

Ops

• Upsert idempotency; delete tombstones for compliance
• Backups, multi-region if needed—eventual consistency semantics per vendor
• Capacity: memory per node vs sharding; replication factor

Cost

• Managed per dimension × count; egress; query units—estimate from peak QPS

Exit condition: Runbook for reindex, scaling, and incident “search degraded.”

Stage 6: Evaluation & Iteration

Goal: Continuous improvement with labeled or proxy eval.

Loop

• Golden query set updated when product changes
• A/B embedding models or rerankers with guardrails on latency
• Monitor click-through, thumbs, or human grading in RAG

Debugging bad retrieval

• Chunk inspection, metadata leaks, wrong tenant filter, stale index

Final Review Checklist

• Metrics and embedding/model versioning plan
• Index family chosen with benchmark evidence
• Hybrid/filter strategy matches product needs
• Ops: upsert, delete, scaling, backup understood
• Eval set and iteration process in place

Tips for Effective Guidance

• Never promise “semantic search understands intent”—ground with eval.
• pgvector vs specialized: trade-offs on scale, ops, features—state honestly.
• Warn: high-cardinality filters + ANN can be slow—design metadata carefully.

Handling Deviations

• Tiny corpus: brute force or simple index may suffice—avoid over-engineering.
• Multimodal: separate embedding spaces or unified model—fusion strategy required.