28 Oct 2026 · Postgres & pgvector

Production RAG in PostgreSQL, pgvector vs Pinecone and Qdrant

When a vector column next to Auth.js is enough, what managed vector DBs actually cost, and how Next.js 15 talks to pgvector through Prisma.

Pinecone is a good product. It is also a second vendor, a second bill, and a second network hop between your Next.js app and the chunks that belong to the same user row in Postgres.

AgenticPacks OSS is Postgres plus Prisma, no vectors: Agentic-Packs/agenticpacks-oss. Pro keeps embeddings in pgvector (1536-dim, cosine <=>) next to Auth.js and Stripe customer ids. The SQL details are in Prisma pgvector migrations.

What you pay for

Figures are order-of-magnitude, USD per month, not a quote.

Setup	At 10k docs	At 1M chunks	Extra moving parts
pgvector on the same VPS as the app	included in the 5 euro node	still on that disk until you vertical scale	extension, raw SQL
Pinecone serverless	tens of dollars	usage + storage	API keys, another region
Qdrant Cloud	similar band	similar band	another cluster
Self-hosted Qdrant	another container	RAM for HNSW	second healthcheck

The self-hosting cost post already argued that serverless glue is the line item you notice. A managed vector DB is the same shape: cheap at hackathon size, loud at SaaS size.

You still pay the embeddings API. pgvector does not make OpenAI free. It removes the retrieval vendor.

When Pinecone still wins

You have hundreds of millions of vectors and a dedicated search team.
You need hybrid search features you do not want to build.
Your Postgres is a tiny managed instance you refuse to grow.

For an AI SaaS with per-tenant docs in the low millions of chunks, Postgres is the boring answer. Filters are WHERE "organizationId" = $1 instead of metadata syntax in a second product.

Next.js shape

Upload or crawl.
Chunk ~500 tokens (Pro's chunker).
Embed in batches. Respect rate limits. See ingestion.
$executeRaw insert as vector(1536).
At query time, embed the question, ORDER BY embedding <=> $q LIMIT k.
Pass chunks into the model. Optionally into LangGraph. See LangGraph streaming.

Authz is a SQL WHERE, not a hope that the vector index is tenant-scoped. Same IDOR lesson as Server Actions.

Ops

The Docker image must be pgvector/pgvector:pg16 (or equivalent). Stock postgres:16 will reject vector. Backups are pg_dump like everything else. Restore includes the extension.

HNSW vs IVFFlat is an index choice, not a reason to leave Postgres. Start IVFFlat after you have data. Rebuild in a maintenance window.

Tired of a $70 vector bill next to a $25 Postgres bill?

AgenticPacks OSS is the relational baseline: Agentic-Packs/agenticpacks-oss.

Pro ships pgvector RAG in the same Compose file. See self-hosted pgvector RAG in Pro.

Decision

If your RAG corpus is tenant documents and your source of truth is already Postgres, add a column. If you are building a standalone search company, buy a search company product.

Where to Go From Here

You do not have to stand up a second database to ship contextual chat.

Option A: Free open-source baseline (MIT)

GitHub: github.com/Agentic-Packs/agenticpacks-oss
What is included: Next.js 15, Compose, PostgreSQL 16, Prisma, Auth.js v5, streaming chat, single-plan Stripe. No pgvector.
Cost: free, MIT.

Option B: Ship with AgenticPacks Pro ($149 one-time)

If retrieval should live next to users and invoices:

Automated Hetzner provisioning: Terraform, Caddy, Compose, pgvector image.
Production multi-agent core: Chunking, cosine search, LangGraph, Redis.
Usage & Unit Economics: Stripe and a token ledger.