Cursor Delivers 99% Faster AI Code Search With Instant Grep

Cursor has introduced a local indexing system for regular expression search, aimed at helping AI coding agents retrieve relevant code faster without scanning entire repositories.

The update comes as coding agents increasingly rely on repeated searches across codebases to gather context while writing and debugging code.

In Cursor’s internal benchmarks, the company said it reduced regex search latency from 16.8 seconds using ripgrep to 13 milliseconds.

The system works by preprocessing codebases into searchable indexes, allowing agents to narrow down a smaller set of candidate files before applying full-regular-expression matching. 

Cursor said this approach addresses a key bottleneck in agent-driven development, particularly in large monorepos where traditional search tools can take several seconds.

Unlike conventional tools such as ripgrep, which scan all files, Cursor’s system filters the search space upfront. “No matter how fast ripgrep can match on the contents of a file, it has one serious limitation: it needs to match on the contents of all files,” the company said.

Cursor said regular expression search remains necessary despite advances in semantic retrieval. “There are specific queries which the model can only resolve by searching with regular expressions,” it added.

The company’s approach builds on trigram-based indexing, in which text is broken into overlapping three-character sequences and mapped via an inverted index. 

It also uses additional techniques, such as probabilistic masks and sparse n-grams, to improve precision and reduce the number of files to be scanned. 

“The goal is to use our index to minimise the amount of potential documents we need to scan,” Cursor said.

A key design decision is to run these indexes locally on the user’s machine rather than on remote servers. Cursor said this reduces latency, ensures results reflect the latest code changes, and avoids the need to sync code externally.

The company added that local execution is important because agents often run multiple searches in parallel while modifying code, making repeated full-repository scans inefficient.