felladrin/MiniSearch

MiniSearch – Private AI‑powered web search you can run yourself

What it does

• A self‑hosted search engine that adds an AI assistant to ordinary web results.
• The search backend is a bundled SearXNG metasearch instance that pulls results from many public search engines.
• Those results are re‑ranked locally with a small cross‑encoder model (ONNX Runtime) and then fed to an AI model that generates a concise answer, citing the sources.
• All AI inference can happen inside your browser (WebGPU or CPU) or via any OpenAI‑compatible API you point it at (Ollama, LM Studio, vLLM, llama.cpp, AI Horde, etc.).
• The UI stores history, caches, and analytics in the browser’s IndexedDB, so nothing leaves your machine unless you enable a remote API.

Why it matters

• Keeps your queries private: no telemetry, no accounts, and you can run the whole stack in a Docker container on your own hardware.
• Gives you the convenience of a modern “search‑with‑answers” experience without relying on third‑party services that see your data.
• Flexible deployment: Docker image, Docker‑Compose, or a Hugging Face Space, plus the ability to set it as your default browser search engine.

How to get started

1. Run the Docker image

   docker run -p 7860:7860 ghcr.io/felladrin/minisearch
   

   Then open http://localhost:7860 in a browser.
2. (Optional) Pin a specific image digest for reproducible builds, or use Docker‑Compose as shown in the README.
3. (Optional) Build from source with the provided docker-compose.production.yml.
4. (Optional) Duplicate the Hugging Face Space to host a cloud‑free instance.

Key technical bits

• Browser AI: Models from 135 M to 4 B parameters are downloaded once and cached; inference runs on WebGPU when available, otherwise on the CPU.
• Reranking: A cross‑encoder model runs via ONNX Runtime inside the container to reorder SearXNG results before they reach the UI.
• Citation‑rich answers: The assistant generates responses that include inline citations to the original search results, and you can request the model’s reasoning or have the answer read aloud.
• Extensible back‑ends: You can point the “AI Processing Location” to a remote OpenAI‑compatible endpoint, letting you use larger models without downloading them locally.

Who might use it

• Privacy‑conscious users who want an answer‑style search without sending data to big tech.
• Developers or researchers needing a sandboxed AI‑augmented search for experiments.
• Teams that want to embed a private search‑assistant into internal tools or intranets.

License

• Apache 2.0 – free to use, modify, and distribute.