Tweet by _chenglou
November 6, 2025
Some thoughts: - Semantic search is nothing new, but the little innovation here is on the data collection side. Cursor agents have metrics for success & failures (e.g. user accepts/rejects), and we can look into those sessions’ searches, and optimize the embeddings so that next release, agents can just directly search better and better - Grep + semantic search is also pretty established; real-world search has always been hybrids (tf-idf and bm25 as pre/post-filter + semantic search) - The real unlock here, in this democratized AI era, is that you can over-include false positive results and have AI trim things down, rather than under-include results (bad) just because they’re too noisy for human reading. Over-including results is basically the premise of e.g. Exa and other AI search engines - “Semantic” is pretty overloaded nowadays; maybe a better term would be “similarity search”. But the ship has sailed
- Author
- _chenglou
- Date
- November 6, 2025
- Canonical URL
- /tweets/chenglou-1986289051226546448-fb1601