mdrag: a Rust CLI to chat with an Obsidian vault

What they built

Orel Lazri's vault grew to snippets, project docs, meeting notes and random thoughts. He wanted to ask questions like 'What were the main challenges in the refactor?' without a cloud service.

He wrote mdrag, a Rust CLI that embeds markdown into SQLite and answers through a local Ollama model, keeping work notes on the machine.

Setup

• Rust with rusqlite + sqlite-vec, ollama-rs, zerocopy
• Embeddings: nomic-embed-text-v1.5 (768-dim)
• LLM: gemma3:1b (temperature 0.2)

How it works

1. Detect each document's primary header level and split on it, then split large chunks on paragraphs
2. Store embeddings in a sqlite-vec virtual table
3. Embed the query, fetch top-5 similar chunks via `embedding MATCH`, and stream an answer from gemma3:1b

Results & numbers

• Initial embedding takes a few seconds; queries are nearly instantaneous
• A 1B model gives good results without expensive hardware

“My notes never leave my machine - this is crucial for vaults that contain sensitive information like work-related notes”— Orel Lazri

Do it yourself

1. Install Ollama and pull nomic-embed-text and gemma3:1b
2. Build mdrag from the author's repo
3. Run `mdrag embed ~/vault` once
4. Ask: `mdrag search "how to debug git ssh issues"`

mdrag embed ~/vault
mdrag search "how to debug git ssh issues"