The Pinecone alternative
that gives you SQL back.

The pattern we see: a team adopts Pinecone for vector search, then discovers the real cost — usage-based bills that grow with traffic, a second data store to keep in sync with their database, and filtering that cannot express what a SQL WHERE clause can. If your source data already lives in Postgres, moving the vectors there too removes a whole system from your architecture.

For most workloads, no. pgvector with an HNSW index on NVMe handles the typical RAG and semantic search scale — hundreds of thousands to tens of millions of vectors — with single-digit-millisecond latency, and we measure ~1,000 QPS at recall@10 0.93 on 250k × 1536-dimension vectors on a $35 node. What you gain is SQL: joins, transactions, full-text hybrid search, and one database to operate instead of two.

At very large scale — billions of vectors, or workloads that are purely vector search with no relational side. Pinecone is built for that and does it well. If that is you, stay. The honest dividing line: if your vectors describe rows that live in a SQL database, pgvector usually wins; if vectors are the whole product at massive scale, a dedicated vector database earns its cost.

Usually substantially, because the pricing models are different in kind: Pinecone meters reads, writes, and storage, so the bill grows with traffic; Rivestack is a flat price per node regardless of query volume. A $15 Solo or $35 Starter node covers what many teams pay Pinecone hundreds for. Send us your vector count, dimensions, and QPS and we will give you a realistic number.

Export vectors and metadata through the Pinecone API, bulk-load into a vector column with COPY, build an HNSW index, and tune ef_search to hit your recall target. Your query path changes from API calls to SQL. For most workloads it is an afternoon of work, and migration help is free — we will plan it with you and pair on the cutover.

Postgres does hybrid search natively: combine vector similarity with full-text (tsvector, pg_trgm) in a single query and rerank with SQL expressions — no sparse-dense vector gymnastics. Because it is all one planner, you can also pre-filter by joins and permissions before the vector scan, which dedicated vector stores struggle to do efficiently.

Your data is encrypted in transit and at rest, backed up daily to separate S3-compatible storage with point-in-time recovery, and HA clusters run Patroni-managed automatic failover. Everything is vanilla PostgreSQL — you can pg_dump your entire database, vectors included, and leave at any moment. No lock-in is the strongest guarantee we can offer, and one Pinecone cannot match.

Yes. The free tier gives you a Postgres database with pgvector in about a minute, no credit card. Load a sample of your vectors, run your real queries with your real filters, and measure recall and latency before you commit. Our live demo at ask.rivestack.io shows pgvector searching 40K chunks of Hacker News if you want to feel the latency first.