The pgvectorscale benchmark everyone is citing this year: 471 QPS at 99% recall on 50 million 1536-dim vectors, competitive with dedicated vector databases. If you’re already running Postgres, that removes an entire second data store from your architecture, and the transactional consistency you get for free is worth more than most teams realize until they’ve debugged a stale-vector bug at 2am. Past 100M vectors, fine, go dedicated. Below that, I don’t have a good reason to add the extra infra anymore.
Related in this Cluster
- NotesWhat Breaks When LLMs Enter Regulated Enterprise Workflows
Most enterprise AI initiatives do not fail on raw model intelligence. They break on the mundane, structural realities…
- NotesThe cost of abstractions
Just spent the afternoon untangling a "helpful" ORM layer that was executing N+1 queries under the hood. Sometimes…
- NotesShadow AI is just a feedback loop nobody built
Talked to a few people this week who admitted to pasting confidential docs into a public chat tool…