Pinecone used to be an essential building block in Aidbase's RAG system.
It would store tens of millions of individual knowledge points from the custom models our users are training on their content.
Our replacement?
A simple Postgres database using pgvector. Hosted on Hetzner.
Here's what changed:
1. Speed: Lower latency on our actual workloads 2. Cost: A fraction of the monthly bill 3. Flexibility: Normal SQL + vectors in one place 4. Control: No vendor lock-in, easier scaling 5. Simplicity: Fewer services, fewer failure points
The myth: You need "specialized" tools for production-grade RAG.
The reality: - You don't need the scale you think. - Optimize where it actually hurts. - Keep ownership of your data.
Simon Høiberg
I just brutally cut another 3rd-party dependency.
Our Vector Database: Pinecone.
Pinecone used to be an essential building block in Aidbase's RAG system.
It would store tens of millions of individual knowledge points from the custom models our users are training on their content.
Our replacement?
A simple Postgres database using pgvector.
Hosted on Hetzner.
Here's what changed:
1. Speed: Lower latency on our actual workloads
2. Cost: A fraction of the monthly bill
3. Flexibility: Normal SQL + vectors in one place
4. Control: No vendor lock-in, easier scaling
5. Simplicity: Fewer services, fewer failure points
The myth:
You need "specialized" tools for production-grade RAG.
The reality:
- You don't need the scale you think.
- Optimize where it actually hurts.
- Keep ownership of your data.
Fancy tools isn't a strategy.
Outcomes are.
In this case:
pgvector won.
2 weeks ago | [YT] | 257