Cursor has developed a method for efficiently indexing large codebases to enhance semantic search capabilities. In a recent evaluation, their approach improved response accuracy by 12.5% and increased user satisfaction. For smaller projects, indexing is nearly instantaneous, but larger repositories can take hours to process if handled inefficiently. By recognizing that many teams use similar codebases, Cursor can securely reuse existing indexes from teammates, reducing time-to-first-query from hours to seconds for large repositories. The initial index is built using a Merkle tree, which identifies changes in files and directories without requiring a complete reprocessing. Each file and folder has a cryptographic hash, enabling quick synchronization. When a change occurs, Cursor breaks the file into chunks and creates embeddings for semantic search. These embeddings are cached, ensuring that unchanged chunks don’t incur additional processing costs during future queries. When a new user joins, they derive a similarity hash from the Merkle tree and upload it to the server. The server searches for matching indexes in a vector database, allowing the new user to benefit from existing indexes without going through the entire indexing process. This is done while ensuring that the client can only access results for code they already have, preventing any data leakage. The entire process significantly speeds up onboarding, especially for larger repositories; for instance, the median time-to-first-query drops from 7.87 seconds to 525 milliseconds.