MicroRAG is a free, open-source AI that answers questions from your own documents, with citations, and never needs the internet. Put your PDFs, notes, scans and recordings on an SD card, plug it in, and ask in plain English. It can also read pictures, turn voice memos into searchable transcripts and check its own answers. Nothing is uploaded. There is no account.
A made-up sample so you can see the shape of an answer. Real answers cite your files.
The real app
The vault as it opens: drop in files, ask a question, and use the tools on the left to see the vault map, find contradictions and check that nothing leaves your computer.
The technical reason
MicroRAG reads a model file of about 2.3 GB every time it starts, then hits the search index on every question. How fast your storage reads decides how fast it feels. That is where a card in a built-in slot wins.
Time to load the 4B model from the same SanDisk Extreme PRO microSD card on the author’s Apple silicon Mac. About 3.5 times faster through the built-in slot. Measured once, so your numbers will differ.
A built-in SD slot has no USB reader or hub in between. In our test, adding a USB reader to the same card made loading 3.5 times slower.
SD cards print performance classes. An A2 card guarantees at least 4,000 random reads per second, and V30 guarantees 30 MB/s sustained writes. Typical USB sticks publish no such floor.
Loading a model is large reads. Searching the index is many small random reads. Those random reads are exactly what the A-class rating measures.
Fair to say: we measured one card through a built-in slot against the same card through a USB reader, not a USB stick head to head. A fast USB 3 SSD drive can match or beat a mediocre card, and Macs without an SD slot need a reader. Buy a card marked A2 or V30 and use the built-in slot if you have one.
Why it exists
Contracts, research, medical notes, client work, family records. Some documents should never be uploaded. MicroRAG keeps the documents, the search index and the AI models on one card, and runs everything on the computer you plug it into.
| MicroRAG | Cloud AI chat | |
|---|---|---|
| Where your files go | Stay on your card and your computer | Uploaded to someone else’s servers |
| Works with no internet | Yes, after the one-time install | No |
| Account or subscription | None. Free and MIT licensed | Usually required |
| Shows its sources | Every answer cites passages, checked against the file | Varies |
| Take it with you | Plug the card into another computer | Tied to your login |
| Raw model power | A small local model. Strong on your documents, weaker than the biggest cloud models on open-ended tasks | Largest models available |
What you get
Everything below runs locally. Nothing here needs a server of ours, because we do not have one.
Ask a question and get an answer built only from your files, with numbered citations. Proof view highlights the exact sentence that backs each claim and shows a strong, partial or weak support badge, plus the page number for PDFs.
Turn on truth and every statement in an answer is checked against the passage it cites. The model has to quote the passage, then the program looks the quote up and checks every number. Each statement gets a tick, a tilde or a cross, and an answer where nothing holds is dimmed. In our test with deliberately broken answers it caught 92% of changed numbers, every invented sentence and 89% of answers to the wrong question, and wrongly flagged 5% of correct ones. It is a second look by a small model, not a guarantee.
Drop in a photo, screenshot, chart, table, receipt or whiteboard. MicroRAG reads what it shows and the text in it, then answers with the picture itself beside the passage that backs the answer. The reading runs on your computer, and the picture model only loads while you add pictures. Handwriting and small print can be misread, so the picture always sits next to the text. Needs the optional pictures pack.
Press record, or drop in a voice memo, call or interview. Whisper turns it into a searchable note with a [mm:ss] time on every paragraph, entirely on your computer, and the audio is not kept. When Whisper marks a change of speaker, each turn gets its own line. One 9.5 minute recording took 34 seconds on one M3 Pro Mac. Download any transcript as a plain .txt file.
A summary, decisions and action items, each with a time that is checked against the transcript. Click a time to read the passage around it. A line whose time cannot be matched to the transcript says so, instead of guessing. Answers about a recording cite the moment it was said.
Qwen3-4B for quality, Qwen3-1.7B for a faster start, or Gemma 3 4B. Switch from a menu inside the app. Only the chat model restarts, your documents stay put, and it remembers your choice. The 1.7B loads in roughly half the time.
One button scans your computer, times a cold read of the card, measures real answer speed and tells you which model it can run. It shows the load time it actually saw, not a guess, and labels anything it could not measure.
Do not take our word for it. A live view lists every connection the app opens. The honest answer should be none.
Scans your vault for passages that disagree, like two contracts with different dates, and shows both sides side by side.
See your whole library as a living map of how documents relate. Spot clusters, gaps and orphans at a glance.
Open your vault from a phone on the same Wi-Fi with a 6-digit PIN. Snap a photo of a page and ask about it. The photo is read on the computer and not kept.
Turn it on and MicroRAG refuses to run while any network is up. Turning it off takes a deliberate step in the terminal.
Drop files onto the app or write a note. Everything is stored as a plain file in your vault folder, so you can back it up or edit it with any tool.
Point it at a folder and get a one-page briefing of what is in there, with sources.
Lock the vault with a passphrase so a lost card is just a lost card.
Photos, screenshots and scanned PDFs are read with on-device OCR (Apple Vision on a Mac). With the pictures pack, a vision model also understands charts, tables and layouts.
The proof moment
Flip the switch. A private AI should not care whether you are online, and this one does not.
Illustration of the idea. In the real app, the Zero-Network Proof view checks this on your own machine, and it says plainly what it can and cannot see.
Quick start
You need internet once, to download the models. After that you never do.
Get MicroRAG from GitHub and put it on an SD card. About 5 GB with the models (about 8 GB with the optional pictures pack), plus room for your files.
It downloads the runtime and the models onto the card. If a download drops, run it again and it resumes.
Open the start file, wait for ONLINE, drop in your documents and ask.
MicroRAG$ cd /Volumes/MicroRAG-SD && sh install.shWho it is for
Search every client file at once without sending any of it to a cloud.
Ask across contracts and spot conflicting terms, with the page cited. Turn a recorded call into minutes with times you can check.
Query your notes, papers and interviews and get answers that point to the source, down to the minute in a recording.
Boats, cabins, flights, field sites, or just a careful household.
Straight talk
Questions
MicroRAG is a free, open-source AI knowledge vault. You put your documents on an SD card, plug it into a computer, and ask questions in plain English. It answers from your own files and cites the passages it used. The documents, the search index and the AI models all stay on the card, and everything runs on your computer.
Yes. You need internet once, to download the runtime and models during install. After that it runs with no connection. The Zero-Network Proof view lists the connections the app makes, and Air-gap mode can refuse to run while any network is up.
Speed. MicroRAG loads a model of about 2.3 GB at every start and searches its index on every question, so storage read speed matters most. On the author’s Mac, the same card loaded the 4B model in 31 seconds through the built-in SD slot and 1 minute 48 seconds through a USB card reader. SD cards also print performance classes, such as A2 for random reads, that typical USB sticks do not publish. We did not test a USB stick head to head, and a fast USB 3 SSD drive can match or beat a mediocre card.
No. There is no account, no cloud and no telemetry. Your files, the search index and the models stay on the card, and questions are answered on your own computer.
An SD card with roughly 5 GB free for the app, the models and speech, or about 8 GB with the optional pictures pack, plus room for your files, and a computer. It was built and tested on a Mac. The Windows and Linux launchers are included but untested. 8 GB of RAM works with the small model, and 16 GB or more is comfortable with the larger one. Reading pictures and Truth mode need 12 GB or more.
Text and Markdown files, CSV, JSON and HTML, Word (.docx) and PDF files. Photos, screenshots and scanned PDFs are read with on-device OCR, which uses Apple Vision on a Mac, and with the optional pictures pack a vision model also describes them and reads charts and tables. Audio files (WAV, MP3, M4A and more) and live recordings are turned into transcripts with Whisper. Quality depends on the scan, the picture and the audio.
Yes. Press record, or drop in an audio file, and Whisper turns it into a note with a time on every paragraph. The recording is processed on your computer and the audio itself is not kept. The transcript is searched like any file, and its briefing is a summary, decisions and action items, each with a time that is checked against the transcript so you can click it and read the passage. The speech model is English only, and it marks a change of speaker but does not say who is speaking.
Yes, with the optional pictures pack (Gemma 3 4B and its vision file, about 3.3 GB). It writes what the picture shows plus the text in it, using Apple Vision OCR to help on a Mac, and answers show the picture beside the passage. The vision model loads only while you add pictures. It can misread small print and handwriting and sometimes adds a small detail that is not there, so check it against the picture.
An optional check on each answer. Every statement is compared with the passage it cites: the model has to quote the passage, the program confirms the quote is really there, and every number in the statement must appear in the passage. Statements are marked backed, partly backed or not backed. In a test with deliberately broken answers it caught 92% of changed numbers, every invented sentence and 89% of answers to the wrong question, and flagged 5% of correct answers by mistake. The test was small, so treat it as a second look, not a guarantee.
Qwen3-4B or Qwen3-1.7B for answers (the app picks by your memory, and you can switch from a menu), nomic embeddings for search, Whisper for speech, and optionally Gemma 3 4B for pictures. Gemma is released under Google’s own terms, not an open-source licence.
Cloud AI chat needs your files uploaded to a server and usually an account. MicroRAG runs locally, needs neither, and shows the source passage for every answer. The trade-off is that it uses a small local model, which is weaker than the largest cloud models on open-ended tasks.
Yes. Any AI can be wrong. MicroRAG cites the passages it used and shows a strong, partial or weak support badge for each answer, so you can check the source before relying on it.
Yes. It is free and released under the MIT license, so you can use, change and share it.
Yes. Phone mode opens your vault on a phone connected to the same Wi-Fi, protected by a 6-digit PIN. Camera Ask lets you photograph a page and ask about it. Phone mode uses plain HTTP on your local network, so use it only on a network you trust.
MicroRAG is developed by William Mapp and released under the MIT license.
Free, open source, and private by design.