Q10: Multi-Tenant Vector DB Architecture: Tenant Isolation, Pre-Filtering & Partition Keys - AI Engineering & LLMs (50 Qs)

1. Enterprise Multi-Tenancy Patterns

In multi-tenant SaaS applications, ensuring that Customer A cannot retrieve Customer B's confidential vector chunks is a critical security and compliance requirement. There are three primary multi-tenant architectural patterns in vector stores:

less

   Pattern 1: Database-per-Tenant         Pattern 2: Namespace-per-Tenant       Pattern 3: Metadata Pre-Filtering
   ┌────────────────────────────────┐    ┌────────────────────────────────┐    ┌────────────────────────────────┐
   │ Tenant A DB   │ Tenant B DB    │    │ Tenant A Index │ Tenant B Index│    │ Shared Global Vector Index     │
   │ [HNSW Graph]  │ [HNSW Graph]   │    │ [Namespace A]  │ [Namespace B] │    │ Chunks tagged: {"tenant": "A"} │
   └────────────────────────────────┘    └────────────────────────────────┘    └────────────────────────────────┘
    Maximum Security • High Cost          Balanced Security • Medium Scale      Low Cost • Strict Filter Logic Req.

2. Multi-Tenant Architectural Comparison

Architecture	Isolation Level	Resource Overhead	Scaling Limit	Best Used In
Database per Tenant	Physical (Dedicated DB process/instance)	Highest (idle memory overhead per tenant)	Hard limit on hundreds of tenants	Highly regulated banking, healthcare (HIPAA)
Collection / Namespace per Tenant	Logical (Separate HNSW graph in shared cluster)	Moderate	Supported up to 10,000 namespaces	B2B enterprise SaaS with distinct tenant boundaries
Metadata Partitioning (Pre-Filtering)	Shared graph with indexed metadata filter	Lowest	Unlimited tenants (millions)	High-volume B2C or multi-user consumer apps

3. Pre-Filtering vs. Post-Filtering Pitfall

[!CAUTION] Never use Post-Filtering for multi-tenancy!

Post-Filtering: Runs top-$K$ vector search across the entire database, then discards items where tenant_id != current_tenant. If Tenant A's documents only constitute 1% of the database, the top 100 vector hits may contain 0 items for Tenant A, returning an empty result!

Pre-Filtering (Iterative Single-Stage): The vector index traverses the HNSW graph while restricting candidate nodes to those matching the tenant bitmap, guaranteeing $K$ results for the target tenant.

python

# Correct Production Pre-Filter Query (Qdrant / Pinecone style)
search_results = vector_client.search(
    collection_name="enterprise_documents",
    query_vector=query_embedding,
    query_filter={
        "must": [
            {"key": "tenant_id", "match": {"value": current_user.tenant_id}},
            {"key": "department", "match": {"value": current_user.department}},
            {"key": "is_confidential", "match": {"value": False}}
        ]
    },
    limit=5
)