what should I use for embeddings if I need cross-tenant isolation?

If you need cross-tenant isolation for embeddings, the safest default is:

Use a separate embedding index / vector store per tenant — and, if the data is sensitive or you need strong guarantees, separate encryption keys and even separate physical databases/accounts per tenant.

Recommended patterns

1. Best isolation: one vector DB/index per tenant

   • Prevents accidental cross-tenant retrieval.
   • Simplifies access control and deletion.
   • Good when tenant count is manageable.

2. Stronger isolation: separate storage + separate keys

   • Each tenant gets its own:
     • embedding index
     • metadata store
     • encryption key
   • This is the usual choice for regulated or high-security workloads.

3. Shared index with tenant_id filters

   • Only use if you need scale/cost efficiency and can tolerate weaker isolation.
   • Must enforce tenant filtering at every query path.
   • Riskier because a bug in filtering can leak data across tenants.

For the embeddings themselves

The embedding model can usually be shared across tenants because embeddings are derived representations, not tenant-specific secrets. What matters for isolation is where you store and query them.

Practical recommendation

• If you’re asking from a security-first standpoint:
  Use per-tenant vector indexes + per-tenant metadata isolation + per-tenant keys.
• If you’re optimizing for cost and scale:
  Shared model, isolated logical partitions, strict authz, and mandatory tenant filters.

Rule of thumb

• High sensitivity / compliance → separate index per tenant
• Moderate sensitivity / many tenants → shared infrastructure, strict logical isolation
• Low sensitivity → shared index with tenant filters may be acceptable

If you want, I can suggest an architecture for:

• Postgres + pgvector
• Pinecone / Weaviate / Qdrant
• multi-tenant SaaS with encryption and key management