How AskFolder searches your folder without uploading a single file

Most tools that answer questions about your documents work the same way: upload the files, let a server read them, send a question to a large language model, get back a written answer. That's fine for a public report. It's not fine for a lease, a medical letter or a year of bank statements.

AskFolder does it the other way round: everything happens on your own computer, inside the browser tab. This post explains how, so you don't have to take it on trust.

Reading the files

When you choose a folder, the browser gives the page access to those files: it never gets anything you didn't choose, and nothing is copied anywhere.

• PDFs are read with Mozilla's PDF.js. A PDF stores words with their position on the page, not in reading order, so AskFolder lays each page out again by where the words sit. A table becomes one line per row, with each cell named by its column header, so a bank statement row reads as "Interest paid: $41.20" instead of a run of numbers.
• Word files (.docx) are read with Mammoth; text, Markdown and CSV files are read as they are.
• Scans and photos have no text at all, only a picture of it. Those pages go through Tesseract, the open-source OCR engine, compiled to WebAssembly. It runs in a background worker, in 11 languages.

Cutting text into passages

An answer should be the paragraph that says it, not a whole 40-page file. So each document is cut into passages of at most about 110 words, following its own structure:

• each heading starts a section, and every passage remembers its heading;
• a section is cut at sentence boundaries, with the last sentence shared with the next passage, so no answer falls into the gap between two;
• a table is kept apart from the prose around it, one passage of its own.

Searching two ways at once

Each question is searched twice, in a background worker so the page never freezes.

• By words: BM25, the ranking classic search engines are built on, plus credit for the exact phrase and for question words standing close together. A word that isn't in the folder is matched to its nearest spelling, so "termintion" still finds "termination".
• By meaning: every passage, and the question, is turned into a vector by all-MiniLM-L6-v2, a small sentence model (23 MB) that runs in the tab through Transformers.js and ONNX Runtime Web. Passages that point the same way as the question match even when no word is shared: "how much notice to end the lease" finds "either party may terminate on sixty days' written notice".

The two lists are then merged by score, each scaled to its own best. Where the words can't tell passages apart (every page of a lease says "lease"), meaning decides; where the words point clearly at one passage, they lead.

The answer is the passage itself, with its file and page, the matching words highlighted. Nothing is generated. There is no chatbot that could make up a clause the document doesn't contain; if no passage says it in those words, AskFolder labels the results as the closest matches instead of pretending.

How you can check that nothing is sent

You shouldn't have to trust a privacy promise. Here's how to check this one in a couple of minutes:

1. Open AskFolder, then your browser's developer tools (F12) and the Network tab.
2. Open a folder and ask a few questions. Every request listed is a file loaded from askfolder.com itself: the page's code, the reading engines, the model. None carries a file or a question, and none goes to another site.
3. Click the page's own request and look at its Headers. The Content-Security-Policy says connect-src 'self' data: and default-src 'self'. That rule is enforced by your browser, not by us: the page cannot open a connection to any other site, even by mistake.
4. Once the page has loaded, switch off your internet and keep asking. It still works, because everything it needs is already on your computer.

askfolder.com is a static site with no server code behind it, so there is also nowhere on our side that could receive an upload. The code is served as written, not minified, and the credits page lists every file and every open-source part with its licence.

How well it works, and what it gets wrong

We test every change against three sets of questions with known right answers. The one that matters most is a set of 39 questions on documents we never tune on: there, the right passage comes first 92% of the time and in the top three 95% of the time, in about half a second per question. On 60 questions about 20 made-up retirees' folders of everyday paperwork, the right passage is always in the top three, and first 9 times out of 10.

What it gets wrong:

• Very short questions ("fees?") give the search little to work with. A few more words help a lot.
• Handwriting is read poorly; Tesseract is made for printed text.
• Very large folders take a while to read the first time, since all the work happens on your computer. Pro can save a folder's reading in the browser so the next visit is instant.
• It finds passages; it doesn't reason across them. "Which of my policies is cheapest?" needs you to compare the answers yourself.

Why there is a paid version at all

AskFolder is free for folders of up to 50 files, with the first 3 scanned pages of each folder read. Pro removes those limits, saves folders in the browser and turns every date in your files into calendar reminders: $9 a month, $79 a year, or $59 once. Since your computer does the work, Pro pays for keeping the readers, the model and the search improving, not for servers.

Even the Pro license is checked without asking anyone: it's a small signed token that the page verifies on your computer with a public key, offline.