# MicroRAG

> MicroRAG is a free, open-source (MIT) private AI knowledge vault that runs from an SD card. It answers questions from the user's own documents with citations, fully offline. Developed by William Mapp.

Site: https://microrag.dev/
Source: https://github.com/Mach-Advisory/MicroRAG

## What it does
- Answers plain-English questions from the user's own PDFs, Word files, notes, text, pictures and recordings, and cites the passages used.
- Runs entirely on the computer the card is plugged into. Documents, search index and AI models stay on the card. No account, no cloud, no telemetry.
- Needs internet once, to download the runtime and models (about 5 GB, or about 8 GB with the optional pictures pack). After that it works offline.

## Features
- Proof view: each citation is checked against its source, with a strong, partial or weak support badge and PDF page numbers.
- Truth mode (optional): each statement in an answer is checked against the passage it cites. The model must quote the passage, the program verifies the quote and every number. In a small test with deliberately broken answers it caught 92% of changed numbers, 100% of invented sentences and 89% of answers to the wrong question, with 5% false alarms. A second look, not a guarantee.
- Pictures: photos, screenshots, charts, tables, receipts and whiteboards are read by a local vision model (Gemma 3 4B, optional pictures pack) and shown beside the passage that backs an answer. Small print and handwriting can be misread.
- Speech to text: record in the app or drop in audio; Whisper (English) makes a searchable transcript with [mm:ss] times, on the computer, and the audio is not kept. Markers show where the speaker changed, not who spoke.
- Meeting minutes: a summary, decisions and action items from a transcript, each time checked against the transcript; click a time to read the passage.
- Choice of models inside the app (Qwen3-4B, Qwen3-1.7B, Gemma 3 4B) and a system check that says which model a computer can run.
- Zero-Network Proof: a live view of the connections the app opens.
- Contradiction Finder: finds passages in the vault that disagree.
- Vault Map: a visual map of how documents relate.
- Phone mode and Camera Ask: use the vault from a phone on the same Wi-Fi with a 6-digit PIN (plain HTTP on the local network).
- Air-gap mode: refuses to run while any network is up.
- Encrypted vault, on-device OCR (Apple Vision on a Mac), instant briefing.

## Facts and limits
- Built and tested on a Mac. Windows and Linux launchers are included but untested.
- Uses small local models (Qwen3 4B or 1.7B, nomic embeddings, Whisper small.en, optionally Gemma 3 4B for pictures), which are weaker than the largest cloud models on open-ended tasks. Answers can be wrong; check the citations.
- Roughly 5 GB for the app and models, about 8 GB with the pictures pack. 8 GB RAM works with the small model; 16 GB or more is comfortable with the larger one; pictures and Truth mode need 12 GB or more.
- License: MIT.

## Why an SD card, not a USB stick
Speed. The app reads a ~2.3 GB model at every start and searches its index on every question. Measured once on the author's Mac, the same card loaded the 4B model in 31 s via the built-in SD slot and 1 min 48 s via a USB card reader. SD cards print performance classes (A2 for random reads, V30 for sustained writes) that typical USB sticks do not publish. A USB stick was not tested head to head; a fast USB 3 SSD can match or beat a mediocre card.

## Compared with cloud AI chat
Cloud AI chat needs files uploaded to a server and usually an account. MicroRAG runs locally and needs neither. The trade-off is model size.