What you will learn
- Know when an in-memory FAISS index stops being enough
- Explain metadata filtering and why it's the feature that matters most
- Compare managed vector DBs (Pinecone, Milvus, pgvector) at a high level
- Combine a similarity search with a metadata filter (hybrid query)
Why this matters
FAISS is perfect for a laptop demo and a single corpus. The moment you have multiple tenants, per-document access rules, or "only search docs from this team after this date," a raw FAISS index falls over. Knowing when and why to graduate to a real vector database is a common system-design interview question.
What FAISS gives you — and what it doesn't
FAISS is a blazing-fast similarity search library. It does one thing: given a query vector, find the nearest stored vectors. It does not give you persistence, multi-tenancy, access control, or filtering by attributes out of the box.
Metadata filtering is the real feature
In production you rarely want "the nearest vectors" — you want "the nearest vectors that this user is allowed to see, from this project, written after this date." That's a metadata filter applied alongside the similarity search:
results = index.query(
vector=embed(query),
top_k=5,
filter={
"tenant_id": user.tenant_id,
"allowed_roles": {"$in": user.roles},
"created_at": {"$gte": "2026-01-01"},
},
)This single capability is why teams move off raw FAISS.
Picking a vector store
- pgvector — you already run Postgres; you want vectors next to your relational data with one fewer system to operate.
- Pinecone — fully managed, you don't want to run infrastructure, metadata filtering built in.
- Milvus / Qdrant / Weaviate — open-source, self-hosted, high scale and tunable indexes.
The right answer is "the one your team can operate." Don't adopt a distributed vector DB to serve 10K documents.
Quick quiz
Q: Your app serves three customer companies from one index, and Company A must never see Company B's documents. What's the minimum capability you need?
How to explain it in an interview
"FAISS is great for a single corpus on one machine, but it's just nearest-neighbor search — no persistence, no multi-tenancy, no filtering. In production I usually need metadata filtering: 'nearest vectors that this user is allowed to see, from this project, after this date.' That's when I reach for a vector database — pgvector if I'm already on Postgres, Pinecone if I want fully managed, Milvus or Qdrant if I'm self-hosting at scale. I pick based on what the team can actually operate, not on benchmarks."