MiniSearch
MiniSearch is a self-hosted, privacy-first search engine with an integrated AI assistant that runs entirely within a single Docker container AI inference can run locally in the browser via WebGPU/CPU (models 135M–4B parameters) or connect to any OpenAI-compatible API backend Search results are aggregated by a bundled SearXNG metasearch instance, reranked locally by a cross-encoder model, and cached in the browser All data—search history, cached results, and conversations—stays in the user's brow
Analysis
TL;DR
- MiniSearch is a self-hosted, privacy-first search engine with an integrated AI assistant that runs entirely within a single Docker container
- AI inference can run locally in the browser via WebGPU/CPU (models 135M–4B parameters) or connect to any OpenAI-compatible API backend
- Search results are aggregated by a bundled SearXNG metasearch instance, reranked locally by a cross-encoder model, and cached in the browser
- All data—search history, cached results, and conversations—stays in the user's browser with zero tracking, telemetry, or third-party exposure
- The system supports configurable retention, activity analytics, browser default-engine integration, Raycast triggering, and Hugging Face Spaces deployment
Why It Matters
MiniSearch addresses a growing demand for private, self-hosted AI search tools that eliminate reliance on third-party APIs and cloud providers, giving users full control over their data and inference pipeline. Its dual-mode architecture—browser-local inference plus optional remote API fallback—demonstrates a practical pattern for balancing privacy, performance, and accessibility in AI-powered search applications.
Technical Details
- Architecture: Single Docker container housing an app server, bundled SearXNG metasearch instance, and an ONNX Runtime cross-encoder reranker; the browser runs a separate UI layer with IndexedDB for local storage
- Inference options: Models from 135M to 4B parameters run in-browser via WebGPU (with CPU fallback); alternatively, users can connect Ollama, LM Studio, vLLM, llama.cpp servers, hosted providers, or the AI Horde via OpenAI-compatible API
- Search pipeline: Queries hit the server, SearXNG aggregates text and image results from multiple engines, the cross-encoder reranks results locally, and responses are cached in the browser
- AI response flow: The assistant reads top reranked results and generates cited answers with conversation memory; model reasoning can be revealed on demand and responses can be read aloud
- Deployment & security: Published as a Docker image with SBOM attestations and provenance; supports digest-pinned builds, access key protection, and Hugging Face Spaces duplication; Apache 2.0 licensed
Industry Insight
- The browser-based WebGPU inference model lowers the barrier to private AI adoption by eliminating the need for dedicated GPU servers, making self-hosted AI search viable for individual researchers and small teams
- The separation of search aggregation (server-side) from inference (browser or remote API) represents a modular architecture that can be adapted for other privacy-sensitive AI applications beyond search
- As regulatory pressure around data privacy increases, tools like MiniSearch that guarantee zero telemetry and local-only storage will become increasingly attractive to enterprise and compliance-conscious users
Disclaimer: The above content is generated by AI and is for reference only.