What formats can a local AI read?
Research does not arrive in one format. It arrives as the folder you actually have, and a local tool reads it where it sits.
The common research formats: PDF, including scanned pages, DOCX, EPUB, HTML, Markdown and plain text. Each is opened by an extractor running on your machine, the text pulled out and indexed, so questions reach every document at once. No format requires an account, a converter service or an upload, because the reading happens where the files already are. And once import finishes, the format stops mattering: every document joins the same searchable, citable index.
The formats that matter.
Born-digital PDFs carry their text and open directly. DOCX files are parsed for their text, headings and tables. EPUB and HTML are text under a wrapper. Markdown and plain text barely need an extractor at all. The odd one out is the scanned PDF, which is a photograph of a page and needs OCR or a vision model to become text; that path is its own subject. Together these cover what a working library actually holds: downloaded papers, reports, exported web pages, lecture notes, drafts, and the long tail of scans.
Why the format list is a privacy list.
Every format a tool claims is also a place it sends your file. When a cloud assistant reads a DOCX, the document is uploaded first; the format support and the data transfer are the same event. With a local tool the two come apart. The extractor runs on your machine, next to the retrieval and the model, so supporting a format costs bandwidth nothing. A DOCX of unpublished results, an EPUB of licensed material, a scan of a confidential record: each is read where it lies, and the copy that exists after import is the one already on your disk.
What happens on import.
Whichever format a file arrives in, the pipeline after extraction is the same. The text is chunked and embedded so retrieval can find its passages, duplicates are detected so re-importing an updated version does not fork the library, and citations keep pointing at the original document rather than the extracted text. Formats stop mattering the moment import finishes: a field notebook's photos, an archive scan and a downloaded paper end up in one index, answerable together.