Pinecone / Weaviate / Qdrant

Dedicated vector databases optimize ANN serving, filtering, and ops at scale. Rough positioning:

| Product | Angle | |---------|-------| | Pinecone | Managed SaaS, minimal ops, serverless tiers | | Weaviate | Open-source + cloud, GraphQL, modular vectorizers | | Qdrant | Rust engine, strong filtering/payload model, self-host or cloud |

Compare on: latency p99 under your k and dimension, metadata filter expressiveness, multi-tenancy, backup/restore, pricing model, hybrid search support, and data residency/compliance.

On interviews: pick a scenario (B2B SaaS RAG, catalog recommendations) and justify managed vs self-hosted vs pgvector—not brand loyalty.

Common pitfalls: choosing by demo UX alone; ignoring egress and embedding API costs; no disaster recovery story; vendor lock-in on proprietary index formats; skipping load test with production filter cardinality.

The trade-off is specialized performance and managed SLAs versus cost, data governance constraints, and another system to secure and observe.

Checklist:

• Map product to ops model: SaaS vs self-host.
• Compare filter + hybrid search capabilities.
• Estimate cost: storage, queries, embedding pipeline.
• Verify backup, migration, and multi-tenant isolation.
• Run load test at target QPS and recall.