Fast, local semantic code search powered by Rust.

Search your codebase using natural language queries like "where do we handle authentication?" — all running locally with no API calls.

• Semantic Search — Natural language queries that understand code meaning
• Hybrid Search — Combines vector similarity + BM25 full-text search with RRF fusion
• Neural Reranking — Optional second-pass reranking with Jina Reranker for higher accuracy
• Smart Chunking — Tree-sitter AST-aware chunking that preserves functions, classes, methods
• Context Windows — Shows surrounding code (3 lines before/after) for better understanding
• Local & Private — All processing happens locally using ONNX models, no data leaves your machine
• Fast — Sub-second search after initial model load, incremental indexing
• Multiple Interfaces — CLI, HTTP server, and MCP server for Claude Code integration

• Installation
• Quick Start
• Command Reference
  • search
  • index
  • serve
  • mcp
  • stats
  • clear
  • list
  • doctor
  • setup
• Global Options
• Search Modes
• MCP Server (Claude Code)
• HTTP Server API
• Database Management
• Supported Languages
• Embedding Models
• Configuration
• How It Works
• Troubleshooting

sudo apt-get update
sudo apt-get install -y build-essential protobuf-compiler libssl-dev pkg-config

sudo dnf install -y gcc protobuf-compiler openssl-devel pkg-config

brew install protobuf openssl pkg-config

# Using winget
winget install -e --id Google.Protobuf

# Or using chocolatey
choco install protoc

# Clone the repository
git clone https://github.com/yxanul/demongrep.git
cd demongrep

# Build release binary
cargo build --release

# The binary is at target/release/demongrep
# Optionally, copy to your PATH:
sudo cp target/release/demongrep /usr/local/bin/

demongrep --version
demongrep doctor  # Check system health

# 1. Navigate to your project
cd /path/to/your/project

# 2. Index the codebase (first time only, ~30-60s for medium projects)
demongrep index

# 3. Search with natural language
demongrep search "where do we handle authentication?"

# 4. Search with better accuracy (slower)
demongrep search "error handling" --rerank

Search the codebase using natural language queries.

demongrep search <QUERY> [OPTIONS]

# Basic search
demongrep search "database connection pooling"

# Show full code content with context
demongrep search "error handling" --content

# Get JSON output for scripting
demongrep search "authentication" --json -m 10

# Search only in src/api directory
demongrep search "validation" --filter-path src/api

# High-accuracy search with reranking
demongrep search "complex algorithm" --rerank

# Quick search with scores
demongrep search "config loading" --scores

# Re-index changed files, then search
demongrep search "new feature" --sync

# File paths only
demongrep search "tests" --compact

Index a codebase for semantic search.

demongrep index [PATH] [OPTIONS]

# Index current directory
demongrep index

# Index a specific project
demongrep index /path/to/project

# Preview files to be indexed
demongrep index --dry-run

# Force complete re-index (delete and rebuild)
demongrep index --force

# Index with a specific model
demongrep index --model jina-code

• All text files respecting .gitignore
• Custom ignore patterns from .demongrepignore or .osgrepignore
• Skips binary files, node_modules/, .git/, etc.

The index is stored in .demongrep.db/ directory inside your project root.

Run an HTTP server with live file watching for continuous indexing.

demongrep serve [PATH] [OPTIONS]

# Start server on default port (4444)
demongrep serve

# Start server on custom port
demongrep serve --port 3333

# Serve a specific project
demongrep serve /path/to/project --port 8080

The server automatically re-indexes files when they change (with 300ms debouncing).

Start an MCP (Model Context Protocol) server for Claude Code integration.

demongrep mcp [PATH]

See MCP Server section for detailed setup.

Show statistics about the indexed database.

demongrep stats [PATH]

📊 Database Statistics
============================================================
💾 Database: /path/to/project/.demongrep.db

Vector Store:
   Total chunks: 731
   Total files: 45
   Indexed: ✅ Yes
   Dimensions: 384

Storage:
   Database size: 12.34 MB
   Avg per chunk: 17.28 KB

Delete the index database.

demongrep clear [PATH] [OPTIONS]

# Clear with confirmation
demongrep clear

# Clear without confirmation (for scripts)
demongrep clear -y

# Clear a specific project's index
demongrep clear /path/to/project -y

List all indexed repositories (searches for .demongrep.db directories).

demongrep list

Check installation health and system requirements.

demongrep doctor

Pre-download embedding models.

demongrep setup [OPTIONS]

These options work with all commands:

demongrep supports three search modes with different accuracy/speed tradeoffs:

Combines vector similarity with BM25 full-text search using Reciprocal Rank Fusion (RRF).

demongrep search "query"

• Speed: ~75ms
• Best for: Most queries, balances semantic understanding with keyword matching

Pure semantic similarity search using embeddings.

demongrep search "query" --vector-only

• Speed: ~72ms
• Best for: Conceptual queries where exact keywords don't matter

Two-stage search: hybrid retrieval followed by cross-encoder reranking.

demongrep search "query" --rerank

• Speed: ~1.8s (adds ~1.7s for reranking)
• Best for: When accuracy matters more than speed

demongrep can act as an MCP server, allowing Claude Code to search your codebase.

1. Build demongrep and note the binary path:

   cargo build --release
   # Binary at: /path/to/demongrep/target/release/demongrep

2. Index your project:

   cd /path/to/your/project
   demongrep index

3. Configure Claude Code

   Edit ~/.config/claude-code/config.json (Linux/Mac) or the appropriate config location:

   {
     "mcpServers": {
       "demongrep": {
         "command": "/absolute/path/to/demongrep",
         "args": ["mcp", "/absolute/path/to/your/project"]
       }
     }
   }

4. Restart Claude Code

Once configured, Claude Code can use commands like:

• "Search for authentication handling"
• "Find all chunks in src/auth.rs"
• "Check if the index is ready"

Request:

curl -X POST http://localhost:4444/search \
  -H "Content-Type: application/json" \
  -d '{
    "query": "authentication",
    "limit": 10
  }'

Response:

{
  "results": [
    {
      "path": "src/auth/handler.rs",
      "start_line": 45,
      "end_line": 67,
      "kind": "Function",
      "content": "pub fn authenticate(...) { ... }",
      "score": 0.89,
      "signature": "fn authenticate(credentials: &Credentials) -> Result<User>"
    }
  ],
  "query": "authentication",
  "total_results": 1
}

Each project has its own index at <project_root>/.demongrep.db/

# Incremental update (only changed files)
demongrep search "query" --sync

# Or explicitly re-index
demongrep index

# Full rebuild (delete and recreate)
demongrep index --force

# Interactive
demongrep clear

# Non-interactive
demongrep clear -y

# Or manually
rm -rf .demongrep.db/

demongrep stats

These languages have full AST-aware chunking that extracts functions, classes, methods, etc.:

These languages are indexed with fallback line-based chunking:

# Index with specific model
demongrep index --model jina-code

# Search must use same model as index
demongrep search "query" --model jina-code

Note: The model used for indexing is saved in metadata. If you search with a different model, you may get poor results. Use --force to re-index with a new model.

Create .demongrepignore in your project root:

# Ignore test fixtures
**/fixtures/**
**/testdata/**

# Ignore generated code
**/generated/**
*.generated.ts

# Ignore large files
*.min.js
*.bundle.js

demongrep also respects .gitignore and .osgrepignore files.

• Walks directory respecting .gitignore and custom ignore files
• Detects language from file extensions
• Skips binary files automatically

• Parses code with tree-sitter (native Rust implementation)
• Extracts semantic units: functions, classes, methods, structs, traits, impls
• Preserves metadata: signatures, docstrings, context breadcrumbs
• Falls back to line-based chunking for unsupported languages

• Uses fastembed with ONNX Runtime (CPU-optimized)
• Batched processing for efficiency
• SHA-256 content hashing for change detection

• arroy for approximate nearest neighbor search
• LMDB for ACID transactions and persistence
• Single .demongrep.db/ directory per project

• Query embedding → Vector search → BM25 search → RRF fusion → (Optional) Reranking

# Index the project first
demongrep index

1. Check if index is stale:

   demongrep search "query" --sync

2. Try different search mode:

   demongrep search "query" --rerank

3. Rebuild index:

   demongrep index --force

If you indexed with one model and search with another:

# Re-index with the model you want to use
demongrep index --force --model minilm-l6-q

# Reduce batch size
DEMONGREP_BATCH_SIZE=32 demongrep index

# Use a different port
demongrep serve --port 5555

# Debug build
cargo build

# Release build
cargo build --release

# Run tests
cargo test

# Format code
cargo fmt

# Lint
cargo clippy

RUST_LOG=demongrep=debug demongrep search "query"
RUST_LOG=demongrep::embed=trace demongrep index

Apache-2.0

Contributions welcome! See TODO.md for planned features.