Serverless vs pod-based, namespaces, metadata filtering, hybrid search.

Pinecone — Managed Vector Database at Scale

Pinecone is a managed vector database optimised for similarity search at scale. Serverless indexes scale to zero and charge per query; pod-based indexes (p1/s1) provide predictable latency. Namespaces partition an index without separate indexes — useful for multi-tenant isolation. Metadata filtering (filter dict on upsert/query) enables pre-filtering before ANN search. Sparse-dense hybrid search combines BM25 sparse vectors with dense embeddings.

Interactive Pipeline Flow

UnderstandImplementOptimise

📁 Index — create / serverless

Create index with dimension matching your embedding model (1536 for text-embedding-3-small). Serverless: pay-per-query, cold start risk. Pod (p1.x1): warm, predictable sub-5ms p99.

Architecture Diagrams

Using namespaces for tenant isolation in Pinecone.

Loading diagram...

Key Design Decisions

🗂️Use namespaces for multi-tenancy

Namespaces are free — they cost no additional storage. Always namespace tenant data: index.upsert(vectors, namespace='tenant-id'). Without it, one tenant's data is visible to all.

🔎Metadata filtering reduces recall — use sparingly

Heavy metadata filtering on a small result set can starve the ANN algorithm of candidates. If filtering to <5% of vectors, consider separate indexes per major filter dimension.

🔀Hybrid search outperforms dense-only for keyword queries

When users search for specific product codes, names, or IDs, dense embeddings often miss exact matches. Sparse-dense hybrid with alpha=0.3 improves precision for keyword-heavy queries.

Discussion