Retrieval Store Selection Guide

Version: 1.0.0 | Last updated: 2026-07-16

Purpose

Select vector infrastructure from measured workload, control, and lifecycle requirements.

Why

Vector count alone does not determine a product or migration threshold. Filtering selectivity, update rate, latency, recall, tenancy, operations, and compliance dominate.

How

Benchmark shortlisted stores with production-shaped vectors, metadata, ACL predicates, churn, concurrency, and failure modes. Gate on filtered retrieval quality, p95/p99 latency, ingest/delete/correction lag, backup/restore, tenant isolation, encryption/key control, observability, portability, and total cost. Start with the existing transactional platform only when it meets gates; select a specialized service when evidence shows a gap.

Tradeoffs

Integrated stores simplify consistency and operations. Specialized stores can improve scale or ANN features but add distributed lifecycle and vendor risk.

Anti-patterns

• Universal cutoffs such as “migrate at N vectors.”
• Selecting from vendor feature matrices without corpus benchmarks.
• Ignoring exact deletion, filtered recall, or restore behavior.

Enterprise Considerations

Review residency, subprocessors, private networking, key ownership, auditability, export format, disaster recovery, and contract exit.

Checklist

• Benchmark uses production corpus, filters, churn, and concurrency.
• Quality, latency, lifecycle lag, cost, and isolation gates are set.
• Restore, deletion, and portability are proven.
• Selection ADR records evidence and re-evaluation triggers.

References

• PostgreSQL pgvector
• ANN Benchmarks methodology (comparative benchmark; reproduce on own workload)
• NIST SP 800-53 Rev. 5

Changelog

• 1.0.0 — 2026-07-16: Replaced vendor/count mandates with workload evidence.