Thoughts

I wrote it down as a choice: the fast distance formula or the exact one. That was wrong. A formula that only shortlists can be as wrong as its speed allows, and choosing between three of them turned out to hinge on a number that runs opposite to throughput.

Object storage is the cheapest durable bytes you can buy and the worst place to search a million vectors. I forked VecPuff and taught its index to keep one bit per dimension resident, so a query pulls a full vector from S3 only when the cheap guess is too close to call.