sonr is a high-performance semantic search tool for local codebases. It consists of a background daemon that manages LLM inference and a CLI for performing fast, context-aware searches.

• Semantic Search: Find code by meaning, not just keywords.
• GPU Accelerated: Uses llama.cpp (via llama-server) for fast embedding and reranking.
• Automatic Model Management: Downloads required GGUF models from Hugging Face automatically.
• Persistent Caching: Embeddings are cached to disk to make subsequent searches near-instant.
• MCP Server: Expose semantic search as a tool via the Model Context Protocol.

• sonr-daemon: A background service that manages two llama-server instances (one for embeddings, one for reranking) and provides a REST API.
• sonr: A lightweight CLI that sends search queries to the daemon and displays results.

• llama-server must be installed and available in your PATH.

cargo install sonr sonr-daemon

sonr-daemon

The daemon will download the default models (Qwen3-0.6B based) and start listening on an available port. It writes this port to a discovery file so the CLI can find it automatically.

sonr "how does the chunking logic work?" ./src

• sonr <query> [paths...]: Search for <query> in the specified paths.
• --limit <N>: Number of results to return (default: 5).
• --json: Output results in JSON format.
• --url <URL>: Connect to a custom daemon URL (overrides automatic discovery).
• --port-file <PATH>: Path to the daemon's port discovery file.

The daemon exposes an MCP (Model Context Protocol) server at /mcp. You can use the CLI as a stdio-to-HTTP bridge:

sonr mcp stdio

This allows MCP clients to use the semantic_search tool with parameters:

• query: Natural language search query
• root_directory: Path to search in
• limit: Maximum number of results (optional)

The daemon supports several flags:

• --port: API port (default: 0 for OS-assigned).
• --port-file: Path to write the assigned port for CLI discovery.
• --embedding-hf-repo: HF repository for the embedding model.
• --reranker-hf-repo: HF repository for the reranker model.
• --gpu-layers: Number of layers to offload to GPU (default: 99).
• --cache-file: Path to persist the embedding cache.

Run sonr-daemon --help for full details.