Vector databases are a solved problem. We don't need a $200/mo managed service when Postgres does it natively.
The Problem
Every AI startup seems to think they need a dedicated vector database. Pinecone, Weaviate, Qdrant — the choices are endless and the marketing is aggressive.
We went with pgvector. Here's why.
Our agents need memory. Kane needs to remember past articles. Midas needs to recall campaign performance patterns. Sentinel needs to match brand mentions against known threats. All of this requires vector similarity search — turn text into embeddings, then find the closest matches.
Why Not Pinecone?
Pinecone is excellent software. It's also $200+/month for production workloads, another service to manage (API keys, monitoring, failover), another vendor dependency (what happens when they raise prices?), and overkill for our scale (we have ~50K vectors, not 50M).
Why pgvector Wins
pgvector is a PostgreSQL extension. It adds vector columns and similarity search to the database you already have. Create a table with a vector(768) column, run a cosine similarity query with ORDER BY, and you're done. No new service. No new SDK. No new bill.
Performance at Our Scale (50K vectors, 768 dimensions)
| Operation | pgvector | Pinecone |
|---|---|---|
| Insert | 2ms | 15ms |
| Query (top 5) | 8ms | 12ms |
| Monthly cost | $0 (included in Cloud SQL) | $200+ |
The Lesson
Boring technology wins. PostgreSQL has been around since 1996. It will be around in 2036. Can Pinecone say the same?
Choose the tool that does the job with the least complexity. Your future self will thank you.
Q: When should you use a dedicated vector database over pgvector?
A: If you're dealing with 10M+ vectors, need sub-millisecond latency, or require specialized features like hybrid search at massive scale, a dedicated solution like Pinecone makes sense. For most startups with <1M vectors, pgvector in your existing Postgres is the pragmatic choice.