From 48 Seconds to 130 Milliseconds: Vector Search in Tinybird (opens in new tab)

A customer needed semantic search over 20 million embeddings. Their first attempt timed out. Three changes turned it into sub-200ms queries. Here's what we learned.