Let’s get this out of the way: Cursor’s semantic search is impressive until you need to find something that isn’t phrased like a human would ask for it. I’ve watched teams burn minutes waiting for the AI to “understand” a query for a specific log format or a unique error code, when a simple `grep -r "0x4A3F" .` would have given them the answer in a second.
The real question isn’t which tool is “better.” It’s about recognizing the context. If you’re exploring a new codebase and you’re thinking, “Where do we handle user session timeouts?”—Cursor’s natural language approach can save you time. But if you’re in the middle of an incident response or a security audit, and you need every instance of a specific cryptographic function or a hard-coded API key pattern, traditional grep (or ripgrep) is still unbeatable. It’s deterministic. It doesn’t guess. It shows you every line, every file, with the exact context you specified.
I’ve lost count of the times a junior engineer has proudly shown me a Cursor search result, only for me to find three critical matches they missed because the AI decided they weren’t “relevant.” For compliance work, that’s a non-starter. You can’t sign off on a GDPR-related code review if your search tool is filtering results based on probabilistic relevance.
So when is it faster to just grep? When you already know what you’re looking for, or at least the exact pattern. When you need completeness over comprehension. And when you can’t afford the overhead of a tool that might helpfully ignore the needle in the haystack because it thinks you meant a different kind of needle.
—Greg
Trust but verify