brodewa369/roowet-semantic-vault-search

This project lets you search a folder of markdown notes, such as an Obsidian vault, by meaning rather than by exact keyword matching. Everything runs on your own computer, so no notes or queries are sent to any outside company, and there are no API costs to search your notes. Under the hood it works as a small pipeline. A background indexer script scans your markdown files, splits them into chunks based on their headers, and sends each chunk to Ollama, a tool for running language models locally, to turn the text into a numeric vector representing its meaning. Those vectors get stored in LanceDB, a local vector database, along with the original text. When you later ask a question, the same process turns your question into a vector and finds the stored chunks whose meaning is closest to it. A file watcher can also keep the index up to date automatically as you edit your notes, and the indexer tracks file hashes so it only re-processes files that actually changed. The main way to use this project is as an MCP server, a small local tool server that AI assistants like Claude Desktop or Claude Code can call directly. Once connected, the assistant gains tools such as search_vault to run a semantic search, read_vault_file to open a specific note, vault_stats for index statistics, and reindex_file to refresh a single file. This means an AI assistant working alongside you can pull relevant context straight from your personal notes vault while answering questions or doing tasks. Setting it up requires Python 3.10 or newer, a running local Ollama installation, and an existing markdown vault to point it at. Configuration happens through a .env file where you set the vault location, the Ollama address, and which embedding model to use, with bge-m3 as the suggested default. The project also ships example identity template files meant to help different AI agents, including non-Claude agents, use the vault consistently. It is released under the MIT license.