Skip to main content

pgvector vs Pinecone: Vectors in Your Postgres or a Managed Service

pgvector keeps vectors in the Postgres you run with no licence fee, while Pinecone is a managed service with plan minimums. Where each runs, what each costs, and why teams moved.

Reading time
4 min read
Sources
14 cited
In this piece

GILD verdict

Choose pgvector when your data lives in Postgres and your vectors fit its index limits, and Pinecone for a hosted store with nothing to run.

Pinecone also takes wider vectors and embeds text server-side, while pgvector carries no licence fee. Teams that moved from Pinecone to pgvector, among them Confident AI and WSO2, cite one database to manage, lower latency and data that stays on their own infrastructure.

pgvector stores vectors in the Postgres you run, free under the PostgreSQL License, while Pinecone is a managed service whose Standard and Enterprise plans carry monthly minimums. Pinecone takes wider vectors and can embed text for you, and pgvector keeps vectors in the same tables, transactions and backups as your data. Teams that moved to pgvector report one less database, lower latency and lower bills.

Choosing between pgvector and Pinecone

Choose pgvector if

  • Your application data already lives in Postgres and you want vectors in the same tables and backups.
  • Your vectors fit an HNSW index at 2,000 dimensions, or 4,000 as halfvec, and you want no licence fee.

Choose Pinecone if

  • You want a hosted service with nothing to run, with embedding generated server-side from your text.
  • Your vectors need up to 20,000 dimensions, or you want every metadata field indexed for filtering by default.

Side by side

DimensionpgvectorPinecone
Software price and plan minimumspgvector is free under the PostgreSQL License, provided that the copyright notice and the license text appear in all copies. It comes preinstalled with Postgres.app and many hosted providers.123Starter is free. Builder is US$20 a month flat. Standard charges a minimum of US$50 a month and Enterprise a minimum of US$500 a month, and usage past the minimum is paid as you go.4
Maximum vector dimensionsThe vector type stores up to 16,000 dimensions. An HNSW index covers vector columns of up to 2,000 dimensions and halfvec columns of up to 4,000.1Dense vectors take up to 20,000 dimensions.5
Where it runs and where the data livespgvector is an extension for Postgres that stores vectors with the rest of your data. You compile it on Linux and Mac for Postgres 13 and later, or on Windows. On Linux and Mac you can also install it with Docker, Homebrew, PGXN, APT, Yum, pkg, APK or conda-forge, and on Windows with Docker or conda-forge. It comes preinstalled with Postgres.app and many hosted providers.1Pinecone runs as a managed service on AWS, GCP and Azure. Starter indexes must be in AWS us-east-1, and Builder, Standard and Enterprise can use any supported region. With Bring Your Own Cloud, which requires the Enterprise plan, vectors, metadata and query and upsert payloads do not leave your own cloud account.6789
Filtering with vector searchWith approximate indexes, pgvector applies filters after the index scan. Iterative index scans scan more of the index when needed.1Pinecone indexes metadata for filtering. All metadata fields are indexed by default.10
Embedding generationpgvector stores and queries vectors your application generates. You can generate them in one language and query in another.1Pinecone can upsert and query with text and generate the vectors server-side.11
Retrieval latency and on-call load after one team left Pinecone (about 40,000 chunks), reported by practitionersReported by practitioners: one team's first-hand account, Sep 13, 2026.The team added the pgvector extension and an embedding column to its existing chunks table for a corpus of about 40,000 chunks. The author reports that p99 latency on the retrieval call dropped by about a third, and puts it down to removing a network hop to a third-party service.12Nguyen Chi Cong reports that the on-call pages for Pinecone rate limits and timeout spikes stopped once the team removed Pinecone.12
Monthly cost one site owner reports (about 200,000 pages, about 2 million vectors), reported by practitionersReported by practitioners: one site owner's first-hand account, Apr 27, 2025. Not a GILD calculation.After moving to pgvector on a US$20 a month Supabase instance, now holding about 2 million vectors, the author reports no fees beyond the base instance cost after a couple of days of setup.13Reddit user fantastiskelars reports that Pinecone serverless cost US$800 to process all the links at first, then US$20 to US$200 a month depending on traffic and crawler activity, for a site of about 200,000 pages and about 15,000 monthly active users.13
Why WSO2 moved its vectors off Pinecone in October 2026, reported by practitionersReported by practitioners: a merged pull request in WSO2's own repository.The change stores the vector data in PostgreSQL with the pgvector extension on WSO2's own infrastructure, and the pull request states that search and indexing behaviour stays the same.14In a pull request merged on Oct 7, 2026, Pathumi Kaveesha writes that WSO2's Smart Search stored WSO2 content in Pinecone, an external, third-party vector database, which raised a data-handling concern.14

What practitioners report, side by side

Each product's figures come from its own practitioners, for the group you pick.

pgvectorThe pgvector project−29 NegativeHow this figure is made
PineconePinecone Systems−63 Strong negativeHow this figure is made
Whose sources to count
Compare pgvector with
  • Pinecone

Overall signal

Teams and leaders rate pgvector −29, 34 points above Pinecone at −63. Most of the gap is in complaints, with 24 of Pinecone's 33 sources negative against 11 of pgvector's 24.

The score counts the authors from teams and leaders that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so pgvector's −29 rests on votes split 4 positive, 7 mixed and 10 negative and Pinecone's −63 on 3, 6 and 23.

A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Run a pilot on your own documents and filters to measure recall against exact search, query latency at your filter selectivity, and memory per million vectors.

Footprint on the AI Application Platform stack

They differ on one of 9 layers. On Tool use, pgvector touches it and Pinecone supports it. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.

Layers where pgvector and Pinecone differ
LayerpgvectorPinecone
Build
Tool useTouchesSupports
The 8 layers where they match
Layers where pgvector and Pinecone match
LayerpgvectorPinecone
Model
Model APIsTouchesTouches
Model gatewayTouchesTouches
Build
OrchestrationTouchesTouches
Data
RetrievalCoreCore
Evaluate
EvaluationTouchesTouches
Operate
ObservabilityTouchesTouches
Cloud deploymentSupportsSupports
GuardrailsTouchesTouches

What practitioners praise

For teams and leaders, neither product has a praise theme with 5 sources from 3 authors.

What practitioners raise as concerns

Among teams and leaders, Filters and the Query Planner and Releases and Host Versions reach 5 sources from 3 authors only for pgvector, at 4 of 5 each. Pinecone needs a pilot on the same retrieval queries before you choose on either. SDK and Tooling Defects and Price and Plan Minimums reach 5 sources from 3 authors only for Pinecone, at 19 of 19 and 6 of 6. pgvector needs a pilot on the same retrieval queries before you choose on either.

What practitioners raise as concerns, Teams and leaders
Raised as a concernpgvectorPinecone
Filters and the Query Planner4 of 5Not a theme in its sources
Releases and Host Versions4 of 5Not a theme in its sources
SDK and Tooling DefectsNot a theme in its sources19 of 19
Price and Plan MinimumsNot a theme in its sources6 of 6

Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.

A count ending in "so far" has under 5 sources from 3 authors among teams and leaders. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.

pgvector SignalsPinecone Signals

14 sources

  1. pgvector on GitHub: READMEOfficial documentationraw.githubusercontent.com
  2. pgvector on GitHub: LICENSEOfficial documentationraw.githubusercontent.com
  3. pgvector on GitHub: META.jsonOfficial documentationraw.githubusercontent.com
  4. Pinecone: PricingOfficial documentationpinecone.io
  5. Pinecone: Operation limitsOfficial documentationdocs.pinecone.io
  6. Pinecone: ArchitectureOfficial documentationdocs.pinecone.io
  7. Pinecone: Bring Your Own CloudOfficial documentationdocs.pinecone.io
  8. Pinecone: Bring Your Own Cloud, deployOfficial documentationdocs.pinecone.io
  9. Pinecone: Object limitsOfficial documentationdocs.pinecone.io
  10. Pinecone: Indexing overviewOfficial documentationdocs.pinecone.io
  11. Pinecone: OverviewOfficial documentationdocs.pinecone.io
  12. Nguyen Chi Cong: we deleted Pinecone from production (Sep 13, 2026)Public sourcechicong115.com
  13. Reddit, fantastiskelars: Pinecone and pgvector costs for a 200,000-page site (Apr 27, 2025)Public sourcereddit.com
  14. wso2-open-operations/pitstop pull request #94: move Smart Search to pgvector (Oct 7, 2026)Public sourcegithub.com
Read the research