What should I use for a secure retrieval layer for sensitive docs?

For a secure retrieval layer for sensitive documents, the best choice is usually a vector-enabled search layer with strong access controls and encryption, not a plain public vector DB.

Good options

• Postgres + pgvector
  Best if you want simplicity, mature security, row-level access control, and easy integration with existing auth.
• OpenSearch / Elasticsearch with security plugins
  Good for hybrid keyword + vector retrieval, especially if you need filtering, auditing, and enterprise controls.
• Managed vector DB with enterprise security features
  Use only if it supports:
  • encryption at rest and in transit
  • tenant isolation
  • fine-grained auth/RBAC
  • audit logs
  • private networking/VPC peering
  • data retention controls

What matters most for sensitive docs

No matter the backend, make sure you have:

• Document-level ACLs enforced at retrieval time
• Encryption in transit and at rest
• Key management via KMS/HSM
• Private networking instead of public endpoints
• Audit logging for all retrievals
• PII/redaction before embedding if needed
• Metadata filtering so users can only retrieve what they’re allowed to see

Recommended architecture

A secure pattern is:

1. Store original docs in a secure object store
2. Split/chunk and embed them
3. Store embeddings + metadata in a private retrieval index
4. At query time, apply auth-based filters first
5. Retrieve only allowed chunks
6. Optionally re-rank and redact before returning results

My practical recommendation

If you’re building this yourself:

• Start with Postgres + pgvector
• Put it behind your existing auth layer
• Use row-level security or strict tenant/user filters
• Keep the raw docs in a locked-down document store

If you need enterprise search across lots of docs:

• OpenSearch is often the better secure retrieval layer

If you want, I can give you:

1. a recommended architecture diagram,
2. a vendor shortlist, or
3. a security checklist for sensitive-doc retrieval.