@rajistics: How does Cursor search 50,000 files in half a second? Two indexes, not one: embeddings for meaning, trigrams for exact text. Trigrams turn search into instant lookups instead of reading every file. A Merkle tree spots which files changed so only those get re-embedded. Turbopuffer holds the vectors 20x cheaper than the old setup. Mostly 1970s and 1990s ideas, combined well.

Rajiv Shah | Explaining AI

Open In TikTok:

Region: US

Thursday 18 June 2026 03:04:20 GMT