pgvector vs Pinecone: vectors in Postgres or a managed vector database
pgvector is a free, PostgreSQL-licensed extension that stores vectors beside relational data with HNSW and IVFFlat indexes, while Pinecone is a managed, serverless vector database with a closed-source core and usage-based pricing, so pgvector fits teams that already run Postgres and want one transactional store, and Pinecone fits teams that want vector search with no database to operate.
Checked
This is the most common build decision in retrieval: keep embeddings in the Postgres database the application already uses, or run a dedicated vector database. The line has blurred in 2026. Pinecone now ships BM25 full-text search and released pg_pinecone, an Apache 2.0 Postgres extension, on 27 October 2026, while the Postgres side gained BM25 through extensions such as pg_textsearch.
For an agent, the questions are practical: whether retrieval needs joins and transactions with application data, how filtered queries behave, what the query loop costs at the agent's real volume, and who operates the database.
What each is
pgvector
A Postgres extension that adds vector, halfvec, bit and sparsevec types, HNSW and IVFFlat indexes, and L2, inner product, cosine, L1, Hamming and Jaccard distances. It inherits Postgres's ACID transactions, point-in-time recovery and joins. Version 0.8.0 added iterative index scans to fix over-filtering.
Pinecone
A managed, serverless vector database with dense, sparse and text fields in one index. Full-text search with BM25 and Lucene syntax reached general availability on 9 September 2026, and bring-your-own-cloud on 23 September 2026.
Pricing model
pgvector is free; the cost is the Postgres host, such as Amazon RDS, Aurora, Supabase or Neon, plus the memory HNSW indexes need as data grows.
Pinecone's Starter tier is free with 2 GB of storage. Builder is $20 a month flat. Standard has a $50 monthly minimum, then $0.33 per GB-month of storage, $4 to $4.50 per million write units and $16 to $18 per million read units. Enterprise has a $500 monthly minimum. Read-unit pricing makes an agent that queries many times per task worth modelling before launch.
Which fits which job
pgvector fits
Applications already on Postgres, retrieval that filters or joins on relational data such as tenant, permissions or dates, and corpora small enough that HNSW memory is affordable.
Pinecone fits
Teams that want no database to operate, large or spiky workloads, and buyers who need bring-your-own-cloud deployment.
In between
pg_pinecone can run its own index inside Postgres, query an existing pgvector HNSW index, or proxy to Pinecone, which gives a migration path in either direction.
How Motiontic approaches it
Motiontic's engineers are vendor-neutral. They load a sample of the client's own documents into the candidate stores, run the real questions through each with the same embedding model and reranker, compare recall, latency and cost per query against an eval set, and hand that eval set over so a later migration is a measured decision.
pgvector and Pinecone, as of 30 October 2026
| pgvector | Pinecone | |
|---|---|---|
| What it is | Postgres extension | Managed, serverless vector database |
| Licence | PostgreSQL License | Proprietary core; pg_pinecone is Apache 2.0 |
| Hosting | Any Postgres host, Postgres 13 or later | Pinecone cloud; bring-your-own-cloud since 23 Sep 2026 |
| Index types | HNSW, IVFFlat | Managed; dense, sparse and text fields in one index |
| Keyword search | Postgres full-text search; BM25 via extensions such as pg_textsearch | BM25 full-text search, GA 9 Sep 2026 |
| Transactions and joins | Yes, inherited from Postgres | No; a separate store kept in sync with application data |
| Pricing model | Free; pay for the Postgres host | Free tier, Builder $20 a month, Standard from $50 a month plus usage |
| Version and activity | 0.8.7 (1 Oct 2026); about 23.5k GitHub stars | Closed source; pg_pinecone released 27 Oct 2026 |
Common questions
Is pgvector free?
Yes. pgvector is released under the PostgreSQL License. The cost is the Postgres host it runs on and the memory its indexes need.
Does Pinecone support keyword or BM25 search?
Yes. Pinecone's full-text search, with BM25 scoring and Lucene query syntax, became generally available on 9 September 2026.
What is pg_pinecone?
An Apache 2.0 Postgres extension Pinecone released on 27 October 2026. It can run its own index inside Postgres, query an existing pgvector HNSW index, or proxy queries to Pinecone.
When does it make sense to move from pgvector to a dedicated vector database?
When index memory, filtered-query latency or load on the transactional database becomes the bottleneck, measured on real traffic. Many agent workloads never reach that point.
Related
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
pgvector vs Qdrant
Vectors inside Postgres compared with an open-source Rust vector engine.
Vector database
Storage and nearest-neighbour search for embeddings, now in many shapes.
HNSW
The layered graph index behind most vector search.
Hybrid search
Keyword ranking and vector similarity, fused into one result list.
Sources
- pgvector README (index types, vector types, Postgres versions)
- pgvector changelog (0.8.7, 2026-10-01)
- PostgreSQL news, pgvector 0.8.0 (iterative index scans)
- Tiger Data, pg_textsearch 1.0 (2026-03-31)
- Pinecone pricing
- Pinecone, full-text search GA (2026-09-09)
- Pinecone, introducing pg_pinecone (2026-10-27)
- Pinecone newsroom (BYOC GA, 2026-09-23)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.