How ChatGPT Reads a PDF
Upload, parse, chunk, retrieve: what actually happens to your document before the first answer.
When you upload a PDF to ChatGPT, the model does not read it cover to cover the way you would. ChatGPT pulls the text out of the file, and for a long document it also stores that text in a private search index. When you ask a question, it finds the passages that look most relevant and hands them to the model along with your question. So for a long PDF, the answer comes from the parts ChatGPT found, not from the whole document.
Here is the part that surprises most people. On every plan except ChatGPT Enterprise, OpenAI's upload help page says ChatGPT reads the digital text in a PDF, not the images inside it. Charts, diagrams and photos are skipped. If ChatGPT has ever ignored the bar chart on page three, that is why.
A few words you will need
- Chunk: a passage of text, often a few hundred words, cut from your document.
- Embedding: a list of numbers that captures what a passage means, so passages about similar things get similar numbers. Embeddings: Meaning as Numbers explains the idea.
- Context window: how much text the model can read at once. Context Windows: The Model's Working Memory covers it.
- Search index (vector store): a private database of your chunks and their embeddings, built so the right passage can be found quickly.
- Retrieval: searching that index and passing the best passages to the model. Why RAG Exists: The Context Problem goes deeper.
One more thing before we start. OpenAI documents parts of this process: the upload limits, how ChatGPT Enterprise handles long documents and images, and the API version that developers build with. It has not published every detail of what happens inside the ChatGPT app. Below, we say which is which.
What happens when you upload a PDF?
Uploading a PDF does not feed it straight to the model. ChatGPT stores the file and processes it into something it can search.
There are several ways in. You can attach a PDF in a chat with the paperclip icon, add it as knowledge to a custom GPT, or add it to a Project. Developers using the API (application programming interface, the way software talks to OpenAI's models) upload files to a vector store and let the model search it with a tool called file_search. In the ChatGPT app, uploaded files can be saved in your Library, separately from the chat that used them, and a saved file stays there until you delete it, according to OpenAI's page on chat and file retention.
Limits apply from the start. As of October 2026, OpenAI's upload help page lists these:
- A document can be up to 512 MB.
- Text and document files are capped at 2 million tokens each. That is about 1.5 million words, since OpenAI's rule of thumb is that 100 tokens are about 75 words.
- Spreadsheets top out around 50 MB, depending on the size of each row, and images at 20 MB.
- You can upload up to 80 files every three hours; Free users get 3 uploads a day.
- Upload storage is capped at 25 GB per user and 100 GB per organization.
- A Project holds up to 5 files on Free, 25 on Go and Plus, and 40 on Edu, Pro, Business and Enterprise. A custom GPT can hold up to 10 knowledge files over its lifetime.
The three numbers below come from that same upload help page.
How does ChatGPT get the text out?
That depends on two things: whether the PDF contains digital text, and which plan you are on.
On every plan except Enterprise (Free, Go, Plus, Pro, Business and Edu), OpenAI says ChatGPT uses text-based retrieval, which "extracts digital text from the file rather than its embedded images". A PDF saved from a word processor has a digital text layer, so this works well. A scanned PDF is different: each page is a picture of text, so there may be little or no digital text to pull out. If you need ChatGPT to work with a scanned document, run it through OCR (optical character recognition, software that turns pictures of text into real text) first, or upload the pages as images, which ChatGPT can look at directly.
This is the pattern OpenAI documents for the API, where a file added to a vector store "will be automatically chunked, embedded, and indexed" (Retrieval guide), and for ChatGPT Enterprise, which sends document text to "a private search index" (guide to file uploads in ChatGPT Enterprise).
ChatGPT Enterprise also reads the images
ChatGPT Enterprise adds a feature called visual retrieval. OpenAI's guide to file uploads in ChatGPT Enterprise describes it: images are pulled out of the PDF and stored together with the text around them, scaled down so they do not take up too much of the context window, and when a question brings up a passage, the linked image can come with it and be read by the model directly.
Which files get this matters. OpenAI's Visual Retrieval with PDFs FAQ says a PDF you upload in a chat, inside a Project conversation, or during a chat with a published GPT gets visual retrieval. A PDF stored as a GPT's knowledge or as a Project File gets text-only retrieval, even on Enterprise.
So the same annual report behaves differently. On Plus, the written pages are searchable and the revenue chart is skipped. On Enterprise, uploaded in a chat, the chart can be found and read too.
- Extracts the digital text
- Skips images inside the PDF
- Searches text passages
- Extracts the digital text
- Extracts images and scales them down
- Links each image to the text around it
- Can hand relevant images to the model
How does ChatGPT search your document?
OpenAI documents the API version in detail, because developers set it up themselves. By default, the API splits text into chunks of 800 tokens, each overlapping the next by 400 tokens so that an idea is not cut in half at a boundary, and developers can change both numbers (Retrieval guide). Search then combines two methods: matching meaning (comparing embeddings) and matching keywords (File search guide). OpenAI's Enterprise guide says ChatGPT Enterprise runs the same kind of combined search.
OpenAI has not published the chunk sizes ChatGPT uses inside the app, so treat the API numbers as a picture of the idea, not as ChatGPT's own settings.
What happens when you ask a question?
When you ask "What are the main findings on page 37?", the model does not open the PDF. ChatGPT searches the index with a phrase based on your question, takes the best matching passages, and puts them in the model's context window next to your question. The model writes its answer from that text.
OpenAI describes one more step for ChatGPT Enterprise: some text goes straight into the context window without any search. Enterprise places up to about 110,000 tokens of document text (roughly 80,000 words) directly in context. If a single document is shorter than that, all of it goes in. If it is longer, the first 110,000 tokens go in and the rest can be reached only through search. With several documents, that space is shared between them. OpenAI publishes these numbers for Enterprise only, says they may change, and does not list the equivalent for other plans, so do not assume the same amount on Free or Plus.
The same guide says some models run one search per question, while reasoning models can run several searches (typically two or three), each one building on what the last one found.
What are the practical limits?
The limits stack up in order.
- File size. A document over 512 MB is refused.
- Text size. Each text or document file is capped at 2 million tokens.
- Direct reading. On Enterprise, about 110,000 tokens can sit in the context window at once. OpenAI's own testing, in its Enterprise file upload guide, suggests about 350 pages of mixed text and images fill that space.
- Search. Everything beyond that has to be found. In the API, a search returns 10 results by default (Retrieval guide). ChatGPT does not publish its number, but the principle is the same: if the passage you need ranks too low, the model never sees it.
This is why specific questions work better than vague ones. "What does the report say about staff turnover in 2025?" gives search a narrow target. "Tell me everything about this document" scatters the search across too many possibilities, so most of the document never reaches the model.
Quick reference: key facts (as of October 2026)
| Property | Value |
|---|---|
| Largest document | 512 MB |
| Most text per document file | 2 million tokens (about 1.5 million words) |
| Text placed straight into context | About 110,000 tokens on ChatGPT Enterprise; not published for other plans |
| Knowledge files per custom GPT | 10 |
| Files per Project | 5 (Free), 25 (Go, Plus), 40 (Edu, Pro, Business, Enterprise) |
| Upload storage | 25 GB per user, 100 GB per organization |
| Images inside PDFs | Read only on ChatGPT Enterprise (visual retrieval) |
| Search type | Keywords plus meaning (documented for Enterprise and the API) |
| Spreadsheets | Usually analysed with Python code (data analysis) |
| Scanned PDFs | Little digital text to extract; run OCR first |
Frequently Asked Questions
Q: Does ChatGPT train on PDFs I upload?
It depends on your plan. On personal plans (Free, Go, Plus and Pro), OpenAI may use your chats and files to improve its models unless you turn off Settings > Data controls > "Improve the model for everyone", or use a Temporary Chat, as its page on how your data is used to improve model performance explains. OpenAI says it does not use content from ChatGPT Business, Enterprise or Edu, or from the API, to train its models by default (Enterprise privacy at OpenAI). Our guide AI Privacy at Work: What You Can and Can't Paste covers the rest.
Q: Can ChatGPT read a scanned PDF without OCR?
Not reliably, outside Enterprise. Text-based retrieval pulls out digital text, and a scanned page is a picture, so there may be nothing to extract. Enterprise visual retrieval reads images inside PDFs with a model that can see, which may help with scanned pages, but OpenAI does not promise how accurately. Running OCR first is the safer route on any plan.
Q: How many pages can ChatGPT actually "see" at once?
It never sees pages as pages; it sees text. On Enterprise, about 110,000 tokens of document text go straight into context (roughly 80,000 words), and OpenAI's testing puts that at about 350 pages for PDFs that mix text and images. OpenAI does not publish a figure for other plans. Past that point, only the passages that search finds are visible.
Q: Why does ChatGPT sometimes miss information that is clearly in the PDF?
Because that information never reached the model. Search may have ranked the passage too low, the information may sit in an image that was skipped, or your wording may not match the way the passage is written. These are search failures more than model failures. A narrower question, or naming the section, usually helps.
Q: Can I force ChatGPT to read the entire PDF instead of retrieving chunks?
There is no switch for it in the ChatGPT app. On Enterprise, a document under about 110,000 tokens goes into context whole. For anything longer, split it into smaller files or paste the section you care about as text, so the part you need is sure to be in front of the model.
What this means for you
You are not handing ChatGPT a document to read. You are handing it a search system with a language model on top. A few habits follow:
- Ask specific questions, and name the section or page when you can.
- Ask for page numbers or quotes, then check them in the PDF.
- If a chart matters and you are not on Enterprise, type its key numbers into your message, or upload a screenshot of the chart as an image.
- Run scanned documents through OCR before uploading.
- For long documents, work one section at a time.
For worked prompts that summarize and draft from documents at work, see AI for Documents and Spreadsheets That Actually Saves Hours.
If you build with the API instead, you control much more: which files go into which vector store, the chunk size and overlap, and how many results come back (Retrieval guide). You can also measure how often search finds the right passages on your own test questions.
Test yourself
A colleague on ChatGPT Plus uploads a 120-page annual report PDF full of charts. She asks "What was the revenue trend in the chart on page 40?" and gets a vague answer. Then she asks "Summarize everything in this report" and gets a summary that skips whole sections. What is going on, and what should she do?
Answer: Two different limits are at work. On Plus, ChatGPT uses text-based retrieval: it extracts the digital text and skips images inside the PDF, so the chart's numbers are invisible unless the text around it states them. She can type the key figures into her message, or upload a screenshot of the chart as an image, which ChatGPT can read. The summary problem is about search. A long report may not all be in front of the model at once, and a question about "everything" gives search no clear target, so only some passages reach the model. She will do better asking about one section at a time ("Summarize the section on staff costs"), asking for page references, and checking them. If her company has ChatGPT Enterprise, a PDF uploaded in a chat there gets visual retrieval, and the chart could be read directly.
For more plain explanations of how AI tools work under the hood, subscribe to Internals Decoded at internalsdecoded.com.
Sources
- Uploading files and audio to ChatGPT (OpenAI)
- Optimizing file uploads in ChatGPT Enterprise (OpenAI)
- Visual Retrieval with PDFs FAQ (OpenAI)
- Chat and file retention in ChatGPT (OpenAI)
- Understanding and counting tokens (OpenAI)
- Retrieval guide, API (OpenAI)
- File search guide, API (OpenAI)
- Data analysis with ChatGPT (OpenAI)
- How your data is used to improve model performance (OpenAI)
- Enterprise privacy at OpenAI
