How Rephlo Reads Your Files

You can give Rephlo a file in several ways — attach it to a chat message, add it to a Knowledge Space, or paste in a screenshot. What happens next depends on what kind of file it is and where you put it, and those two questions are easy to mix up.

This page answers one thing: I gave Rephlo a file — what did it actually do with it?

The short version

Every file you hand to Rephlo travels one of three roads:

Road	When it's taken	What Rephlo does
Text	The file already contains text — a normal PDF, Word document, Markdown, CSV, code	Reads the text straight out of the file
OCR	The file is a picture of text — a scanned PDF, a photo of a page, a fax	Recognizes the text on your computer, then treats it like the Text road
Vision	The file is a real image — a chart, a screenshot, a photo	Sends the picture itself to a model that can see

The difference between the first two matters more than it sounds. A PDF from a scanner looks identical to a PDF from Word when you open it, but to software one is text and the other is a photograph of text. Before Rephlo could do OCR, a scanned PDF was simply an unreadable picture — there was nothing to search and nothing to quote.

The second question: attachment or Space?

The road decides how the file is read. Where you put it decides how long it lasts and how it's used.

	Attach to a message	Add to a Knowledge Space
Lifetime	This conversation only	Stored until you delete it
Reuse	Re-attach it every time	Available to every chat and command that uses the Space
How it's used	Placed into the prompt for that message	Split into chunks and indexed, so Rephlo retrieves only the relevant parts
Good for	A one-off question about one document	Reference material you'll come back to — a handbook, a contract set, research notes

Both go down the same three roads. A scanned PDF is OCR'd whether you attach it or add it to a Space; the difference is what happens to the text afterwards.

Why Spaces don't just send the whole document. A large document won't fit in a model's context window, and stuffing it in would be slow and expensive. Spaces index the text so Rephlo can pull out the handful of passages that actually answer your question. See RAG Retrieval & Tuning.

Putting both questions together

What you give Rephlo	Where you put it	What happens
PDF with a text layer, Word, Markdown, CSV, code	Attached to a message	Text is read and included in that message
The same file	Added to a Space	Text is read, chunked, and indexed for retrieval
Scanned PDF (no text layer)	Attached to a message	Text is recognized on your device, then included
Scanned PDF	Added to a Space	Text is recognized on your device, then chunked and indexed
Screenshot, chart, photo	Attached to a message	The picture goes to a vision-capable model
Screenshot, chart, photo	Added to a Space	Any text found in it is indexed; the picture itself isn't retrievable

Reading scanned documents (OCR)

OCR — optical character recognition — is the technology that turns a picture of text back into text. Rephlo does this entirely on your own computer. The document is never uploaded anywhere to be recognized.

Turning it on and off

Go to Settings ▸ Privacy ▸ Read scanned documents on this device.

It is on by default. With it off, Rephlo skips recognition and sends the file straight to the AI model instead — which is a reasonable choice if you're using a model that reads documents well on its own and you'd rather not spend the local processing time.

The one-time download

The first time OCR runs, Rephlo downloads a small text-recognition model — about 15 MB — and tells you when it's ready. This happens once, in the background. If the download fails (usually a connection problem), Rephlo says so; scanned documents won't be read until it succeeds.

Choosing the document language

Under the same setting you'll find Scanned document language. It defaults to Auto, which follows your app language, so most people never need to change it.

Change it when the documents you scan are written in a different language than the app interface. This matters more than it might seem: recognizing Japanese or Korean pages with a Latin-alphabet recognizer produces unusable text rather than slightly-worse text.

Recognition is tuned for these languages:

Languages
English, French, Spanish, German, Portuguese	Shared Latin recognizer
Japanese, Chinese (Simplified)
Korean
Vietnamese

Other languages still work on a best-effort basis, but accuracy isn't guaranteed for them.

Very long documents: the "Partial OCR" badge

Recognizing a page takes real work on your machine, so indexing is capped at 200 pages per document. If a scanned document runs past that limit, Rephlo indexes whatever text the file already had and marks it in the Space with a Partial OCR badge.

The badge exists so you're never misled. If you see it, some pages are missing from search results for that document — you're looking at part of it rather than silently missing the rest. For a very long scanned document, splitting it into smaller files is the practical fix.

Images

Images work differently from documents, because an image can carry meaning that no amount of text extraction captures — a chart's shape, a diagram's layout, a photograph's subject.

If your model can see

Vision-capable models get the picture itself. Rephlo does not OCR it, and does not replace it with text. This is deliberate: text extraction would throw away exactly the visual understanding you attached the image for.

If your model can't see

Not every model can process images — on-device models can't, and neither can many text-only cloud models. Previously, attaching an image to one of these meant the image was quietly dropped and you got a confident answer from a model that never saw your picture.

Now, when the model can't see, Rephlo reads any text out of the image on your device and passes that along instead. If the image contains no readable text at all — a photo of a person, say — you get a visible notice rather than silence, so the model can tell you it couldn't see the image instead of inventing one.

This is a fallback, not an upgrade. Recognized text from a chart is not the same as understanding the chart. If visual detail matters, use a vision-capable model.

Images and privacy redaction

If you've turned on Remove personal info before cloud requests, images are covered too.

Rephlo scans an attached image with on-device OCR before it goes to a cloud model. If it finds personal information, the image is replaced by its redacted text. If it finds nothing to redact, the image is sent as-is.

Be clear about the limits of this. OCR reads text; it cannot catch faces, signatures, or handwriting it fails to read. For images where you can't accept that risk, turn on Block send if redaction uncertain — Rephlo will hold back images it can't read rather than send them.

See Privacy & Data for the full redaction settings.

Quick answers

Question	Answer
I added a scanned PDF and search finds nothing	Check Settings ▸ Privacy — is "Read scanned documents on this device" on, and did the one-time model download finish?
The recognized text is gibberish	The document's language probably doesn't match. Set Scanned document language explicitly.
My document shows a "Partial OCR" badge	It's longer than the 200-page indexing limit. Split it into smaller files.
I attached an image and the model ignored it	The model likely can't see images. Rephlo now passes along any text it finds; for visual understanding, switch to a vision-capable model.
Does my scanned document get uploaded for OCR?	No. Recognition happens on your computer.
Can I use OCR with an on-device model?	Yes. OCR is independent of which model you use — it runs locally either way.

Related pages

• Spaces — what a Knowledge Space is and how it stores documents
• Managing Spaces — adding and removing documents
• RAG Retrieval & Tuning — how Rephlo picks which passages to use
• Privacy & Data — redaction, local processing, and what stays on your device
• On-Device vs Cloud Models — which model to use, and why