Pinecone vs pgvector: which fits your stack?

What does a vector database do?

It stores embeddings and finds the ones closest to a query. An embedding is a list of numbers produced by a model, and the distance between two embeddings measures how related the two pieces of content are.^[1] A search for "cancel a contract" can then find a passage about "terminating the agreement" even though the words differ.

Searching every vector gives exact results but slows down as data grows. Approximate nearest neighbor (ANN) indexes trade a little recall for speed. pgvector states this directly: by default it runs exact search with perfect recall, and adding an index switches to approximate search.^[2]

Pinecone and pgvector are two common ways to get this. One is a managed service built for vectors; the other adds vectors to PostgreSQL.

How does pgvector work?

pgvector is an open-source extension that lets you store vectors with the rest of your data in Postgres.^[2] It adds:

• Types: vector, halfvec (half precision), bit (binary) and sparsevec.^[2]

• Distance operators: L2, inner product, cosine and L1, plus Hamming and Jaccard for binary vectors.^[2]

• Two index types: HNSW has better query performance in its speed and recall tradeoff, but builds more slowly and uses more memory. IVFFlat builds faster and uses less memory, with lower query performance.^[2]

Index limits matter when you pick an embedding model. The vector type can be indexed up to 2,000 dimensions, halfvec up to 4,000 and bit up to 64,000; sparsevec indexes take up to 1,000 non-zero elements.^[2] OpenAI's text-embedding-3-large returns 3,072 dimensions by default, so it needs halfvec or a shortened output to be indexed.^[1]

Because it is Postgres, you keep ACID transactions, point-in-time recovery and joins.^[2] An embedding and the row it describes can change in one transaction.

How does Pinecone work?

Pinecone runs as a managed service on AWS, GCP and Azure.^[3] You do not run servers or tune indexes. Its documentation describes the serverless design:^[3]

• Storage: vector data sits in distributed object storage.

• Separate paths: writes and reads take different paths that scale on their own, so queries do not slow writes and writes do not slow queries.

• Namespaces: each index is split into namespaces, and every read and write targets one. Pinecone suggests a namespace per customer when you need to isolate tenants.^[4]

• Indexing: small slabs of data use fast, light indexing; as they grow, the system merges them into larger slabs with more sophisticated methods.

Pricing is usage-based. The Standard and Enterprise plans bill database storage, read units and write units, among other items, against a monthly minimum, and there are a free Starter plan and a flat-fee Builder plan.^[5] Check the vendor's page for current rates.

How do they compare side by side?

How does filtering differ?

Both filter on metadata, but the mechanics differ.

In pgvector, filters on an approximate index are applied after the index is scanned. The README's example: if a condition matches 10% of rows, with HNSW and the default ef_search of 40, only 4 rows will match on average.^[2] Version 0.8.0 added iterative index scans, which keep scanning until enough results are found, in a strict or relaxed order.^[2] For filters on a few distinct values, the README suggests partial indexes; for many values, partitioning.^[2]

Pinecone takes filter expressions with each query. Operators cover equality and comparison, $in and $nin lists, $exists, and $and, $or and $not.^[6]

What about hybrid search?

Semantic search can miss exact strings such as product codes. Both tools let you add a keyword signal.

pgvector points to Postgres full-text search for hybrid search, and its sparsevec type can hold sparse embeddings.^[2] Pinecone describes sparse vectors as capturing keyword information.^[4] Its hybrid guide offers one index holding dense and sparse vectors, or separate indexes whose results you merge.^[7] Our guide to hybrid search with dense and sparse embeddings covers how to weight the two.

Which should you choose?

Start from your data and your team, not from benchmarks.

• pgvector fits when the vectors describe rows you already hold in Postgres, you need transactions or joins across them, your team runs Postgres well, and your data must stay on infrastructure you control.

• Pinecone fits when you want no database operations for vectors, need per-tenant namespaces, or expect your vector workload to grow apart from your application database.

Other options exist, including other managed vector services and search engines with vector support. The same questions apply to each. Run a pilot on your own content with real queries and real filters, and measure recall and latency on both before you commit.

The store matters less than what goes into it. Our guide to enterprise RAG architecture covers the data pipeline, and advanced RAG reranking and chunking covers retrieval quality. Our data infrastructure for AI service covers this choice as part of our AI consulting work.