Frequently asked questions

Open DevTools with F12, go to the Network tab, clear the list and filter by Fetch/XHR, then drop a file into the workspace. No request carries the file: parsing, chunking and embedding happen in the same tab that displays the result. The only outbound requests are GET calls for model weight files, which carry a file name and nothing else. Switch DevTools to offline mode after a model is cached and the tool keeps working.

In your own browser profile. Text chunks and their vectors sit in IndexedDB for this site, the model weights in Cache Storage, and your settings in local storage. Nothing is written to a server, so there is no remote copy for us to hold, leak or hand over. Clearing site data for this site removes all three at once.

PDF with a text layer, DOCX, TXT, Markdown, CSV, HTML and JSON. The importer sniffs the magic bytes first, so a file renamed from .doc to .docx is rejected rather than half-parsed. PDF page numbers and DOCX heading levels survive parsing, which is why a citation can point at a page or a section instead of a character offset.

Encrypted PDFs, scanned PDFs with no text layer, legacy .doc files, and any single file above 25 MB. Each rejection is reported by its own name — PDF_ENCRYPTED, PDF_NO_TEXT_LAYER, LEGACY_DOC, FILE_TOO_LARGE, UNSUPPORTED_FORMAT — instead of a generic failure, so you know whether to remove a password, run OCR, convert the format, or split the file. A PDF that yields fewer than 40 characters of text is treated as a scan rather than as an empty document.

Yes, with a smaller allowance: the per-knowledge-base limit drops from 40 files to 10 on mobile. Embedding runs several times slower on a phone CPU than on a laptop, and the optional generator is usually too slow to be pleasant there. On a phone, extraction mode with citations is the practical choice.

The default embedder is bge-small-zh-v1.5 at about 25 MB for Chinese or all-MiniLM-L6-v2 at about 23 MB for English; multilingual-e5-small, for mixed Chinese and English text, is about 120 MB. The optional generator, Qwen2.5 0.5B to 1.5B in 4-bit quantisation, costs 400 MB to 1.0 GB and downloads only after you press a button that names the size.

Nothing was precomputed for you. Chunks are embedded on your hardware, and with the local generator enabled every token is produced there too, at roughly 3-8 tokens per second on CPU/WASM. A hosted service amortises one GPU across many users; here you spend your own CPU time and get the privacy in return. The second question about the same document is faster, because the index and the model are already local.

By Google AdSense, on every page, up to six labelled slots (three in-page banners, two side rails, one collapsible anchor). Personalized ads load personalized only after you agree. Decline and the tool behaves identically: outside the EEA, UK and Switzerland you still get non-personalized ads, and inside them no ad script loads at all. There is no paid tier, no account and no data product.

No. The models here run inference only: they compute vectors and answers, never gradients, and no feedback path writes your text back into a model. We could not train on your content even if we wanted to, because no copy of it reaches us at any point.

Yes. The terms place no fee and no licence restriction on the tool: it is free to use, including for work and commercial purposes, and your documents, questions and answers remain yours. Two things sit outside that answer. The downloaded models carry their own licences, listed per model on the models page and governed by the upstream model page rather than by us. And no accuracy is warranted, so a decision that has to satisfy a regulator, an auditor or a client still needs a human to check the cited passage against the source document.

Up to 40 files, 200 MB in total and 20,000 chunks per knowledge base, with a 25 MB ceiling on any single file; on phones the file cap is 10. When a limit is reached the importer says which one it was. Removing the largest file you no longer need clears the block in most cases.

After the model weights are cached, yes. Put the browser into offline mode, close the tab, reopen the site and ask a question: parsing, retrieval and extraction all run locally. The model download is the only step that needs a network, and it happens once per model, not once per session.

A session exports from the chat panel as Markdown, JSON or plain text, and the export carries the model name, the strictness setting and every citation, which is the most reliable way to keep an answer next to its sources. Answers also copy to the clipboard, and browser printing produces a PDF of any page. No part of the export touches a server, because there is no server to run the job.

Read the citations before doubting the tool. If the passage that contains the answer is not among them, the problem is retrieval rather than reading: raise strictness, ask a narrower question, or re-import the document when repeated headers and footers are crowding out the body text. If the right passage is cited and the answer still misreads it, switch strictness to “Documents only” and read the passage yourself.

Only as text. In a PDF a table is flattened into lines in reading order, so a wide table with merged cells loses its column alignment; a CSV keeps each row intact. Where the columns carry the meaning, converting the table to CSV before importing gives noticeably better answers.

Scanned PDFs with no text layer are not supported: the parser finds no characters and returns nothing, and the importer says so rather than pretending the file was indexed. No OCR model is bundled, because one would add hundreds of megabytes to the download for a case a browser handles poorly. Run OCR in another tool first, then import the text-layer PDF or the exported TXT.

No. Encrypted PDFs are rejected before parsing and no password field appears, because the app has no decryption path. Remove the password in a PDF reader and import the unprotected copy; the original file on your disk is untouched either way.

A hosted assistant usually answers better: larger models, faster replies, and abilities this tool does not attempt, such as reading charts. The trade is what happens to your file. If the documents are public and you want the strongest answers, a hosted assistant wins. If the text is confidential, regulated, or simply not yours to upload, keeping it local is the point.

Usually yes, and it is easier to approve than a cloud tool: the site is static files, the single outbound request is a model download, and you can host the model weights on an internal mirror and point the app there instead of at a public host. Two caveats: the machine needs a current browser with WebAssembly support and writable browser storage, and managed devices that lock down site storage will not run it at all.

Current Chrome and Edge, plus recent Firefox and Safari versions with WebAssembly. WebGPU, used only for the optional generation model, is present in current Chromium browsers; everywhere else the pipeline falls back to CPU/WASM, which is slower. A private window works, but its index is discarded when the window closes.

Ad slots sit outside the chat workspace, and no document text, chunk or question is handed to the advertising script. What an ad request carries is the page address plus the standard device and browser details any ad request includes. If you would rather have no third-party script at all, decline the consent message — inside the EEA, UK and Switzerland that stops every ad request outright; elsewhere it narrows requests to non-personalized ones.

Yes. Add PDFs, Word files, spreadsheets and notes together and they are searched as one collection; each result says which file and which paragraph it came from.

You can add as many as you like. Strictly speaking there is no upload step: the browser opens the files for us straight from your disk, so they never leave your computer.

Exact search returns only sentences containing your text as typed - whole words in English, case-insensitive - with no model involved, so it works the moment the page loads, before anything is downloaded. Fuzzy search takes the semantic path and returns sentences that mean something similar, which is what you want when you are hunting a concept rather than a phrase.

No. Open the page and use it: no account, no installer, no sign-in. Once the model is cached it keeps working with the network off.

Two differences that matter: your file is never sent anywhere, and every answer is assembled from the sentences in your own documents with numbers that jump back to the paragraph, so you verify the source rather than a rewrite.

Exact search walks every sentence of every file and lists all of them in document order, unfiltered, instead of handing you a few “most relevant” hits. Only fuzzy search ranks, and the two are separate buttons.

Both. Chinese matches as a substring, English as whole words, case-insensitively, and a single file can mix the two.

The first pass parses and indexes, so the bigger the document the longer that step takes, and the progress bar tells you where it is. Searching an index that already exists takes milliseconds, and the index lives in your browser, so you do not rebuild it next time.

The index is kept in your browser’s local storage, so it is still there when you come back. Clearing your browser data removes it along with everything else - that is the “deleting really deletes” part.