pgvector vs Pinecone

Quick verdict

pgvector is an open source extension for PostgreSQL 13 and later, under the PostgreSQL License. You enable it with CREATE EXTENSION vector and get a vector column type (plus halfvec, bit and sparsevec), distance operators, exact search by default and approximate search with HNSW or IVFFlat indexes. Vectors are ordinary columns, so they share PostgreSQL transactions, joins, replication, backups and roles. Version 0.8.7 was released on 1 October 2026, and most managed PostgreSQL services ship it, often a few versions behind (the Amazon RDS extension list showed 0.8.2).

Pinecone is a proprietary managed vector database. You create indexes through its API, upsert records with vectors and metadata, and query by similarity; Pinecone runs the infrastructure on AWS, Google Cloud or Azure, and Enterprise customers can use BYOC in their own cloud account. Document indexes, generally available since 2 September 2026, combine dense vectors, sparse vectors and BM25 full-text fields. Pinecone also hosts embedding and reranking models.

Side by side

Key differences

One database or two

With pgvector, embeddings are columns in the same tables as your products, documents or users. An insert of a row and its embedding is one transaction, deletes are consistent, and a query can join vectors to prices, permissions or stock in plain SQL. Backups, replication, roles and monitoring are what you already run for PostgreSQL.

-- PostgreSQL with pgvector: semantic search filtered by relational data
CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE documents (
    id         bigserial PRIMARY KEY,
    tenant_id  int NOT NULL,
    body       text NOT NULL,
    embedding  vector(1536)
);
CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops);

SELECT id, body
FROM documents
WHERE tenant_id = 42
ORDER BY embedding <=> $1
LIMIT 10;

With Pinecone, the vectors live in a separate service. Your application or a pipeline must write each record to both your primary database and Pinecone, keep them in step on updates and deletes, and copy any metadata you want to filter on (up to 40 KB per record). In return you get a vector-specific API, serverless scaling and hosted models, and your PostgreSQL server is not loaded by vector queries.

Index control and filtered queries

pgvector gives you the index choice. HNSW builds a graph and needs no training data; IVFFlat divides vectors into lists and should be built after the table has data. Index limits are 2,000 dimensions for vector and 4,000 for halfvec, so very large embeddings need half precision, binary quantization or dimension reduction. The README warns that approximate indexes apply WHERE filters after the index scan, so a selective filter can return fewer rows than requested; iterative index scans (since 0.8.0), partial indexes and partitioning are the documented remedies.

Pinecone does not expose ANN algorithm choices or tuning parameters on serverless indexes. Filters are applied as part of the query against metadata, and namespaces split data by tenant before the search. You trade control for not having to tune.

Hybrid search and embeddings

pgvector does not do hybrid search by itself. The README suggests combining it with PostgreSQL full-text search (tsvector, plainto_tsquery, ts_rank_cd) and merging the two result lists with Reciprocal Rank Fusion or a cross-encoder, which you write in SQL or application code. It also has no embedding generation: your application calls a model and stores the vector.

Pinecone builds both in. A document index can hold dense vectors, sparse vectors and BM25 full-text fields, so one service handles keyword and semantic signals, and Pinecone-hosted embedding and reranking models can be called at upsert and query time.

Scaling and operations

pgvector scales with PostgreSQL. The README lists vertical scaling (memory, CPU, storage), read replicas, and sharding with tools such as Citus or PgDog; HNSW indexes work best when they fit in memory, so large collections mean a larger server. On a managed service, the provider handles patching and backups, but you still choose instance size, and the pgvector version is whatever the provider ships.

Pinecone serverless indexes have no capacity to size. Paid plans add dedicated read nodes for high-throughput indexes, and namespace ceilings reach 100,000 per index on Standard and 1,000,000 on Enterprise. Older pod-based indexes are legacy: new Standard and Enterprise organisations since 18 August 2025 cannot create them.

Pricing and licensing

pgvector is free under the PostgreSQL License. Your cost is the PostgreSQL server that runs it: your own hardware, or a managed service such as Amazon RDS, Aurora, Cloud SQL, AlloyDB, Azure Database for PostgreSQL, Supabase or Neon, each with its own pricing. If you already run PostgreSQL, the marginal cost is the extra memory, storage and CPU for vectors and their indexes.

Pinecone (pinecone.io/pricing, 7 November 2026, $ sign, currency code not stated): Starter free with up to 2 GB storage, 2M write units and 1M read units a month on AWS us-east-1; Builder a flat $20 a month with up to 10 GB storage; Standard a $50 monthly minimum, then read units $16 to $18 per million, write units $4 to $4.50 per million, storage $0.33 per GB per month and egress $0.10 per GB after 100 GB, with a 3-week trial including $300 of credits; Enterprise a $500 monthly minimum with higher unit rates and a 99.95% SLA; BYOC by sales.

Pricing checked on the vendors' official pages on 7 November 2026. Prices change; confirm before buying.

Where each one leads

Limitations

When to choose each

When neither is right

• You want a dedicated vector database but need it to be open source and self-hostable: see Pinecone vs Weaviate and Milvus vs Qdrant.
• You want vector search inside a full-text search engine: see Elasticsearch vs OpenSearch vector search.
• You only need embedded search in one application or notebook: see Chroma vs LanceDB.
• You are still choosing a relational database: start with MySQL vs PostgreSQL or MongoDB vs PostgreSQL.

Final recommendation

Frequently asked questions

Sources

• pgvector on GitHub (README)
• pgvector CHANGELOG
• pgvector LICENSE
• Amazon RDS for PostgreSQL extension versions
• Microsoft Learn: pgvector on Azure Database for PostgreSQL
• Pinecone pricing
• Pinecone indexing overview
• Pinecone 2026 release notes

Checked November 2026.

How we research comparisons: our editorial method.