Uploading files and what the AI can read

Attach PDFs, Word, PowerPoint, Excel, text and audio files — not just pictures — and learn exactly how much of each file your AI can read back, page by page.


Kaizendex keeps your files, not just your photos. A PDF, a Word or PowerPoint document, an Excel sheet, a plain text / Markdown / CSV / JSON file, or an audio recording can be attached to any entry the same way a picture can — and once it's in, it's read, remembered by where things are inside it, and searchable.

What you can attach

KindFilesSize limit
PicturesPNG, JPEG, WebP, GIF, SVG10 MB (SVG 2 MB)
VideoMP4, WebM50 MB
DocumentsPDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), .txt, .md, .csv, .json25 MB
AudioMP3, M4A, WAV, OGG, WebM50 MB

The file keeps its original name — "Lease 2026.pdf" stays "Lease 2026.pdf" — so you can find it by name later. Other types (zip archives, installers, unknown formats) are refused at the picker with a message saying why, rather than uploaded and then ignored.

What happens after you upload

Within about a minute the file is processed in the background — nothing waits on it, your entry saves and syncs as usual:

  1. A short description is written (the "what is this" line — see Files get a description you can edit).
  2. The file's text is pulled out — a PDF's pages, a document's paragraphs, a spreadsheet's rows, a recording's transcript, the words in a photo.
  3. That text is stored as pieces that remember where they came from: page 7, sheet "Q3" rows 12–40, 14:20–14:50 of the recording, paragraphs 4–9.
  4. The language of the text is detected, and a coverage record says exactly how much of the file was read.

Every one of those facts — status, language, coverage, which model, when — is stored on the file itself, so nothing about what the AI knows of your file is hidden.

How much of a file is read

FileWhat is readWhat is not (yet)
PDF with real textEvery page
Scanned PDF (pictures of pages)Up to 30 scanned pages per run, page by pagePages past the 30th — the file is marked partial and says which pages are still unread
Word / PowerPointEvery paragraph / every slidePictures inside the document; speaker notes
ExcelEvery non-empty row of every sheetFormulas (their results are kept), styles, charts
Text, Markdown, CSV, JSONThe whole file
AudioThe whole recording, transcribed, in 30-second stretchesWho was speaking (only when the transcriber can tell)
VideoThe cover frame is describedThe soundtrack is not transcribed yet — the file is marked partial
PhotosAny legible text in the picture

A file whose reading stopped short is marked partial, and the note beside it says why. A file that could not be read at all is marked failed with the reason — and your original file is untouched either way.

What the AI can read back

When you ask the assistant about a file, it has a tool that reads the file itself, piece by piece — the page, the rows, the minute of the recording — and it always knows whether it has read all of it. It is told, plainly, never to present a slice as the whole file: if a PDF is 80 pages and it has read 30, it says so. Reading an entry's fields and notes is a different, lighter tool that does not open attached files — so "read this entry" never silently means "read the PDF on it".

Search works the same way: a result that comes from inside a file points at the exact place — page 7 — and the assistant can jump straight there.

Fixing the text yourself

If the extracted text is wrong — a scan read badly, a name misspelled in a transcript — you can correct it. Your text then becomes the truth: the assistant reads your version from then on, search updates, and re-processing the file later will not overwrite your correction unless you explicitly ask it to.

You can also ask for a file to be processed again at any time (after a better reader ships, say). The original bytes are always kept, so nothing is ever lost by trying again.

Both live in Files & Photos: every file is a row there, with its status, language, coverage, text and more as columns you can choose, sort and filter — and the row's ⋯ menu holds Read text (to correct it) and Reprocess file.