ragrep 🔍

Warning

WIP big time. This codebase is full of broken glass, sharp edges, and dragons.

A semantic code search tool that uses embeddings to find similar code snippets across your codebase. Use it to search for things like:

• "how do we handle http request errors?"
• "where is our data validation?"

Features

• Semantic search - Find code by meaning, not just keywords
• Fully local - No API keys, no cloud dependencies
• Fast server mode - Keep models loaded for 10x faster queries (0.5s vs 7s)
• Auto-reindex - File changes trigger instant reindexing (~200ms)
• Multi-language - Rust, Python, JavaScript, TypeScript via tree-sitter
• Smart caching - Reuse embeddings for unchanged code chunks

Installation

cargo install ragrep

Building from Source

Prerequisites

• Rust toolchain (1.75.0 or later recommended)
• SQLite 3.x

git clone https://github.com/yourusername/ragrep.git
cd ragrep
cargo build --release

The binary will be available at target/release/ragrep

Quick Start

Important

First run downloads models (~1.5GB) to ~/.cache/ragrep/models/

1. Index Your Codebase

ragrep index

2. Start the Server (Recommended)

# Start server in background
ragrep serve &

# Server loads models once (4.6s)
# Queries now run in 0.5s instead of 7s
# File edits auto-reindex in ~200ms

3. Search

ragrep "handle http request error"

Usage Modes

Server Mode (Fast, Recommended)

# Terminal 1: Start server
$ ragrep serve
[INFO] Loading embedder model...
[INFO] Loading reranker model...
[INFO] File watcher started
[INFO] Server listening on .ragrep/ragrep.sock

# Terminal 2: Search (uses server automatically)
$ ragrep "error handling"  # 0.5s ⚡

# Edit files - auto-reindexed!
$ vim src/main.rs  # Save triggers reindex (~200ms)

Standalone Mode (Fallback)

# Works without server (loads models each time)
$ ragrep "search query"  # 7s 🐌

Auto-Reindexing

When server is running:

• Watches .rs, .py, .js, .ts files
• Respects .gitignore and .ragrepignore
• Debounced (default 1000ms)
• Smart caching reuses embeddings for unchanged chunks
• Only git repositories (gracefully disabled otherwise)

Configuration in .ragrep/config.toml:

[server.git_watch]
enabled = true
debounce_ms = 1000

Supported Languages

• Rust (.rs)
• Python (.py)
• JavaScript (.js)
• TypeScript (.ts)

More languages can be added via tree-sitter parsers.

How It Works

Indexing:

• Scan files (respects .gitignore and .ragrepignore)
• Parse with tree-sitter into semantic chunks (functions, classes, etc.)
• Generate 1024-dim embeddings (mixedbread-ai/mxbai-embed-large-v1)
• Store in SQLite with sqlite-vec extension

Searching:

• Embed query → vector similarity search → rerank with BAAI/bge-reranker-base
• Results show file path, line numbers, and relevant code

Auto-Reindexing:

• Watch source files for changes
• Incremental reindex with smart embedding cache
• Only re-embed modified chunks (10-15x faster than full reindex)

Development

See DEVELOPING.md for architecture, setup, and contribution guidelines.

License

This project is licensed under the MIT License - see the LICENSE file for details.