Why AI Can't Read Your PDF, and How to Fix It

When AI can’t read your PDF in Insulin, check the model the turn ran on before the file: Anthropic, OpenAI and Gemini models read a PDF itself, layout, images and charts included, while the Fours-hosted models and the open-source providers get only its text, up to 50,000 characters, and a scanned PDF has no text to give.

---

You attach a PDF to a chat and ask about it. The reply says it couldn’t read the file. Or it answers from the text and misses the chart that carries the point. Or it read a report just like this one yesterday, and today it can’t.

In Insulin, two things decide how much of an attachment reaches the model: the file, meaning its format and size, and the model the turn runs on. You set which model runs first, once per agent, and when a turn switches to another, the reply says so. Below are six symptoms with their causes and fixes, from Fours’ documentation of what each model receives, then the limits that apply and when a knowledge base is the better home.

Why can’t AI read your PDF?

AI can’t read your PDF when the model the turn ran on receives less of the file than your question needs. In Insulin, Anthropic, OpenAI and Gemini models read a PDF itself, “layout, images and charts included.” The Fours-hosted models and the open-source providers, OpenRouter, Baseten, DeepInfra, Fireworks AI, Together AI, Cloudflare Workers AI and DigitalOcean GradientAI, get only a PDF’s text, up to 50,000 characters.

A scanned PDF is a PDF made of page images, with no text layer behind them. On the text-only path it has nothing to give, so “the model is told it could not read the file rather than left to guess.”

The turn runs on	From a PDF, it gets	From an image, it gets
An Anthropic, OpenAI or Gemini model	The PDF itself, layout, images and charts included	The image
The Fours-hosted GLM 5.3 Flash or Gemma 4 31B	Only the text, up to 50,000 characters	The image
Any other Fours-hosted model, or a model from one of the open-source providers	Only the text, up to 50,000 characters	A note that an image is there that it can’t view

For a PDF, the model decides, not the file, so the same PDF can be read as a document on one turn and as bare text on the next. The groups go by who serves a model: anything from one of the open-source providers is on the text-only path. The models Fours hosts are what most agents can run on before you connect anything, and the fallback a turn degrades to when your own model fails.

Which symptom do you have, and how do you fix it?

Match what you see to a row: the first two rows are about the file, and the last four are about the model the turn ran on. Each fix changes either the file or the agent’s Default model, the model it runs first.

#	What you see	Why	What to do
1	The agent says it can’t read the file, and names the formats it can	The file isn’t a PNG, JPG, GIF or WebP image, a PDF, a Word (.docx, .doc), Excel (.xlsx, .xls) or PowerPoint (.pptx) file, or a text file. It uploads anyway	Export it to one of those formats, such as a HEIC photo to JPG or a legacy .ppt deck to .pptx, and attach the copy
2	The image uploaded, but the agent can’t read it	It’s over 5 MB: “a larger image uploads, but the model can’t read it”	Resize, compress or crop it to under 5 MB
3	The agent says it couldn’t read a scanned PDF	The turn ran on a Fours-hosted model or an open-source provider, which gets only a PDF’s text, and a scan has none	Run the turn on an Anthropic, OpenAI or Gemini model, or OCR the scan so the PDF carries text
4	The answer misses a chart, an image or the layout, or ignores the end of a long PDF	The same path: only the text, up to 50,000 characters	Make an Anthropic, OpenAI or Gemini model the Default model; for a long PDF, attach only the pages you’re asking about
5	The agent says an image is there that it can’t view	The turn ran on a Fours-hosted model other than GLM 5.3 Flash and Gemma 4 31B, or on an open-source provider	Make an Anthropic, OpenAI or Gemini model the Default model, or, among the Fours-hosted models, GLM 5.3 Flash or Gemma 4 31B
6	A file like this worked yesterday; today it doesn’t, and the reply ends in an italic note	The turn failed over from the Default model to one that gets less of the file	Fix what the note says went wrong with the first model, then ask again

OCR is optical character recognition: it turns a scan’s page images into a text layer, which gives the text-only path something to read.

Rows 1 and 2 need no settings. Because an unreadable file “still uploads,” an attachment chip in the composer is no proof the model read it; the reply is. Rows 3 to 6 come down to two questions: which model ran, and which one should.

Which model read your file, and how do you change it?

The Default model runs first, and a reply that switched to another model ends with an italic note naming both. Insulin “selects the model for you”: there’s no per-conversation picker, and you choose the model once per agent, not per message. The Default tells you where a turn started; the note, if there is one, tells you where it ended.

Agent	Where its Default model is	Who can change it	Whose providers it can use
The built-in Insulin assistant	The pencil in the conversation header, labelled Insulin model settings: the row with the Default badge	You: click another model’s row	Yours, connected under Settings → Integrations
Your Personal agent	Its Default model field	You	Yours
An Organization agent	Its Default model field	Its owner, an ADMIN or an EDITOR on it; anyone below Editor sees the settings read-only	The organization’s

So a fix in rows 3 to 5 has two halves:

1. Make an Anthropic, OpenAI or Gemini model available to the agent. For the built-in assistant and your Personal agents, connect your own key under Settings → Integrations; they “never reach the organization’s keys.” An Organization agent draws on the providers the organization has connected.
2. Make it the Default, then ask again. On a shared agent you can only use, ask its owner or an Editor to change it.

An Organization agent’s Default model is one setting for everyone who uses it, so an agent your team sends PDFs to should start on a model that reads them.

Why did a file that worked yesterday stop working?

Because the turn failed over: the Default model hit a usage limit or was unavailable, and the turn continued on a model that gets less of the file. Insulin fails over automatically when a model call fails, through “a provider outage, a usage limit, or a rejected key.” Your own connected models are preferred, and the Fours-hosted ones are “the fallback the turn degrades to,” so a turn that starts on an Anthropic model can end on one that gets only the PDF’s text.

The switch isn’t silent. The reply ends with a short italic note naming both models:

first-model hit its usage limit — continued with second-model.

It reads was unavailable instead when an error caused the switch. The wording tells you what to fix:

• If the first model hit its usage limit, its quota is spent: ask again once the limit resets, or raise it with that provider.
• If it was unavailable, a provider outage or a rejected key caused the switch: check that provider’s connection under Settings → Integrations, or give an outage time.

Who picks the model, and what happens when it fails explains why Insulin fails over instead of stopping, and what that trades.

What limits apply to an attachment?

Insulin takes up to 10 files per message, each up to 25 MB, and its models read images up to 5 MB; a model that reads a PDF itself is also held to its provider’s limits. Insulin’s size limits apply to each file, while the providers’ count the whole request.

Limit	Value	Set by
Files per message	Up to 10	Insulin
Size per file	Up to 25 MB	Insulin
Images a model can read	PNG, JPG, GIF and WebP, up to 5 MB each	Insulin
PDF text on the text-only path	Up to 50,000 characters	Insulin
Request size	32 MB	Anthropic
Pages per request	600, or 100 when the request’s context window is under 1M tokens	Anthropic
PDF format	A standard PDF, with no password or encryption	Anthropic
File size	Each file under 50 MB, and 50 MB combined per request	OpenAI
Page images	Read only by a model with vision capabilities	OpenAI

Both Anthropic limits count the entire request. They apply “on the entire request payload, including any other content sent alongside PDFs,” so several long PDFs in one message can add up to more than a provider takes, well inside Insulin’s limits. A PDF with “many small-font pages, complex tables, or heavy graphics” can fill the context window before it reaches the page limit, and Anthropic’s advice is to split it into sections. Remove a password before you attach a PDF, too: Anthropic takes only a standard PDF.

Reading a PDF itself also costs more than reading its text. Anthropic converts each page to an image and extracts its text, so a page counts as text tokens, typically 1,500 to 3,000 depending on density, plus the image; OpenAI notes that sending both “can increase token usage.” For a long document where only a few pages matter, send those pages.

When does a knowledge base fit better than an attachment?

When the document is text you’ll ask about in many conversations, or long enough that each question needs only part of it. A knowledge base splits each document into chunks and indexes them for search, and an agent searching it “cites the source document for each result,” so a question reaches the passage it needs wherever that passage sits in the document.

	Attach to the chat	Add to a knowledge base
Formats	Images, PDFs, and Word, Excel, PowerPoint and text files	PDF, DOCX and text formats including .txt, .md, .csv, .tsv, .log, .json, .html, .xml and .yaml
Size	Up to 25 MB a file and 10 files a message; images read up to 5 MB	Up to 25 MB a file; pasted text up to 2 MB
What the model gets	Depends on the model the turn runs on	The passages search returns, each with its source document
Where it reaches	This conversation, whose Files tab keeps it	Every conversation with an agent it’s attached to; an agent searches up to three knowledge bases per turn
Best for	A one-off question; a chart or a scan, on a model that reads the PDF itself	A long text document you’ll keep asking about

Two cautions before you move a document across:

• A knowledge base doesn’t fix a scan. Its documentation describes no text recognition, so OCR a scanned PDF before you add it.
• Not every attachment can go in. Images, Excel workbooks and PowerPoint decks aren’t knowledge-base formats: export a sheet to CSV, and keep a screenshot or a deck as a chat attachment.

If an agent then can’t find the document, why an AI knowledge base isn’t finding documents works through the checks in order, from Test search to indexing status.

Frequently asked questions

Why can’t AI read my scanned PDF?

On a Fours-hosted model or an open-source provider, the model gets only a PDF’s text, and a scan has none, so it says it couldn’t read the file. Run the turn on an Anthropic, OpenAI or Gemini model, which reads the PDF itself, or OCR the scan first.

Which models read a PDF’s charts and images in Insulin?

Anthropic, OpenAI and Gemini models, which read a PDF itself, layout, images and charts included. The Fours-hosted models and the open-source providers get only a PDF’s text, up to 50,000 characters, so charts, images and scanned pages don’t reach them.

Why can’t the agent see my screenshot?

The image may be over 5 MB, or not a PNG, JPG, GIF or WebP, or the turn ran on a model that can’t view images. Of the Fours-hosted models and the open-source providers, only the Fours-hosted GLM 5.3 Flash and Gemma 4 31B see images.

How many files can I attach to an Insulin chat?

Up to 10 files per message, each up to 25 MB. Images are read up to 5 MB each. A file outside the supported formats still uploads, but the agent is told it can’t read it, along with the formats it can.

Why did my PDF work yesterday but not today?

Check the end of the reply. If an italic note names two models, the turn failed over from the Default model, after a usage limit, an outage or a rejected key, to one that may get only the PDF’s text. Fix the first model, then ask again.

Should a long PDF go in a knowledge base instead?

If it’s text you’ll ask about again, yes. A knowledge base splits each document into chunks for search and cites the source document for each result. It takes PDF, DOCX and text formats up to 25 MB per file; OCR a scanned PDF before adding it.

Takeaways

• What reaches the model depends on the model. Anthropic, OpenAI and Gemini models read a PDF itself; the Fours-hosted models and the open-source providers get its text, up to 50,000 characters.
• A scanned PDF needs a model that reads PDFs, or OCR. On the text-only path it has nothing to give, and the model says so rather than guessing.
• The Default model runs first. Set it once per agent; on a shared Organization agent, its owner, an ADMIN or an EDITOR sets it.
• An italic note at the end of a reply means the turn failed over. Fix the first model it names, and the next turn starts there again.
• Insulin’s size limits are per file; a provider’s count the whole request. For a long text document you’ll come back to, a knowledge base is the better home.

A symptom table explains why a file went unread; whether it belonged in the chat at all is the earlier decision. For that one, see attach a file to chat, or add it to a knowledge base.