Vector database

• You run RAG, semantic search, or similarity-heavy features at scale.
• You need permissioning, TTLs, and metadata filters on embeddings.
• Latency is critical and brute-force search is too slow.

Vector DB choice affects UX speed, cost per query, and security. PMs must ensure tenancy, PII handling, and observability are in place. Query costs can dominate unit economics; good filters and sharding reduce spend. The database also constrains recall quality if the wrong distance metric or indexing strategy is chosen.

Start with hybrid search (dense + sparse) plus strong metadata filters. Enforce namespaces per customer and TTL for sensitive data. Monitor p95 latency, recall@k, and cost per 1k queries. In 2026, prefer managed services with built-in encryption, observability, and disaster recovery unless you have infra talent in-house.

Migrating to a managed vector DB with filtered ANN dropped p95 search time from 420 ms to 95 ms and cut cloud spend 22%, while keeping recall@5 steady at 0.82 on product FAQs.

• Treating the vector DB as a dump without metadata hygiene.
• Choosing cosine vs. dot without validating on your data, degrading relevance.
• Ignoring multi-tenancy isolation, risking cross-customer leaks.