A vector database, built in the open

What it is, where it came from, and why it's shaped the way it is.

Vector Panda mascot, a panda in an orange hoodie and beanie, reading a book with a chart card and a stack of books beside it, lightbulb idea over its head, gazing toward the copy on the left

Vector Panda started as an itch from a previous project, one that added 100,000 vectorized images to S3 every week with no good way to actually query or index them. What I wanted was “S3 for vectors, with Standard and Infrequent Access tiering.” Nothing on the market really fit.

So in June 2025 the first commit was a Rust filesystem watcher: detect a parquet file, log a line, that was the whole feature. A year of work later it's a clustered vector database that auto-optimizes its own indexing on your real data and charges by the GB instead of the query. The blog has the receipts: every false start, every renamed service, every “felt wrong immediately” decision that got walked back two weeks later.

The product is shaped by what felt wrong about other vector databases: the configuration tax, the per-query bills, the approximate algorithms that quietly miss results. The pricing is shaped by the fact that there's no VC pressure here: one operator, self-hosted infrastructure, one monthly bill. That's why hot/warm/paused are storage-only and why queries are free. It's also why this is still private beta. I want to be sure it works for the niche I'm building for (teams who've outgrown self-hosted but don't want hosted-vendor bills) before I open the door wider. The roadmap is long. Plenty more to come.

Vector Panda mascot in glasses and orange hoodie, smiling

Reach out if you want to pitch in, try it out, share some ideas, or just tell me my music is bad and I should feel bad. :)