Freelance, Independent
Grimore-MD: local-first RAG for your notes that even runs on Android
I built Grimore-MD, an open-source (MIT) tool that turns a vault of notes and documents into something you can query. Everything runs locally: no API keys, no cloud.
What it does:
- Watches your vault (Markdown, PDF, EPUB, DOCX and more) and auto-tags, summarizes and categorizes every file with a local LLM via Ollama
- Hybrid search: BM25 + embeddings merged with Reciprocal Rank Fusion, all inside SQLite. No separate vector DB
- Answers questions with citations back to the exact note, page or heading
- Ships a stdio MCP server, so Claude Desktop, Cursor or Zed can query your vault
Since it touches personal files, the defaults are paranoid on purpose: dry-run on the first scan, LLM calls locked to localhost, git snapshots before any write, and binary files are never modified (sidecar notes instead).
Fun part: I build it mostly from my phone in Termux, and it runs on Android too.
Next up is an eval harness (Hit@1, Hit@3, MRR@10) to measure retrieval quality on real personal vaults. Question for the room: how do you evaluate RAG on messy personal data with no ground truth? Synthetic Q&A, hand-labeled sets, something else?
RAGLocal First
MCPOpen Source
OllamaLocal-First






Comments
0 comments