Vibe-Coded Local Search: Ask Your Files Questions by Voice or Photo

What it does

It’s a local-first desktop search tool for Windows, macOS, and Linux that indexes videos, audio, Office docs, PDFs, and Markdown files. You can query by typing, speaking (up to 30 seconds), or uploading an image, and it uses a hybrid of Faiss vector search and Whoosh full-text search to retrieve results. The backend runs Python and FastAPI with local models like BGE-M3 and CN-CLIP, while the frontend is an Electron-Vue shell styled after Notion.

The interesting bit

The project wears its “vibe coding” origin openly—every prompt and dev document is included in the repo for study—and it doubles as an MCP server and Agent Skill provider, meaning Claude Desktop, Cursor, or VS Code can call it as a native tool to search your disk. That turns a standalone utility into a plug-in long-term memory for other AI assistants.

Key highlights

• Multimodal input: voice (via FasterWhisper), text, and image (via CN-CLIP) queries against local content.
• Hybrid retrieval: combines Faiss vector similarity with Whoosh keyword search, plus an expandable terminology library that the README claims boosts recall by 60% in specialized domains.
• Privacy toggle: runs fully offline by default, with optional cloud LLM and embedding APIs (OpenAI, DeepSeek, Alibaba) if you want heavier models.
• Desktop integration: packaged as an Electron app with a Notion-inspired design system, system font stack, and Lucide icons.
• AI agent plumbing: exposes an MCP server and standardized Agent Skills so external coding assistants can query your filesystem directly.

Caveats

• Setup is involved: the easy path is a Windows-only bundled package; macOS and Linux require manual developer deployment, including separate downloads for Ollama, FFmpeg, and several gigabytes of models hosted on Baidu Netdisk.
• The license is split: free for non-commercial use and modification, but commercial deployment requires explicit authorization.
• The documentation is Chinese-first; an English README exists, but quick-start links and community chat point to Chinese services.

Verdict

Worth a look if you want a hackable, local semantic search layer that can also serve as long-term memory for your AI coding tools. Skip it if you need a zero-config, cross-platform binary or a permissive commercial license out of the box.