Multi-tenancy

Isolating data, cost and limits between customers in a shared AI platform.

Stable: this rarely changes. Reviewed 17 Oct 2026.

Key ideas

• Tenant ID on every record, vector, trace and cache key.
• Choose pooled (shared index with filters), bridge (namespace per tenant) or silo (dedicated resources) per tier.
• Per-tenant token budgets and rate limits; attribute cost from traces.
• Never share prompt or semantic caches across tenants unless content is public.

Common mistakes

• Cross-tenant leakage through a shared semantic cache or a missing retrieval filter.