Uploads and their limits

The upload button accepted your file. That tells you almost nothing. It means the file was small enough to store. Whether the AI will read all of it, some of it, or a summary of it is decided later, out of sight. This lesson is about where the limits are, what happens when you hit one, and how to feed a long document to an AI so nothing important gets dropped.

Three gates, in order

Gate one: file size. Measured in megabytes. This is the only limit that gives you a clear error. Too big, and the upload fails on the spot. It is also the least interesting limit, because a 50-page PDF of plain text is tiny, while a 3-page PDF full of photos can be huge. Size measures pictures, not words.

Gate two: pages or tokens per file. Each tool caps how much of one file it will convert into text. Past that cap, the rest of the file is either refused or ignored. Sometimes you get a message. Sometimes you do not.

Gate three: the context window. This is the one from AI Basics lesson 6: the model's short-term memory, measured in tokens. Everything has to fit in it at once: the hidden system prompt, your file, your question, the whole conversation so far, and the answer being written. Today's top models hold about a million tokens, roughly 700,000 words, but many apps use less than the full window, and a long chat eats into it. When the file plus the conversation outgrows the window, something has to give.

What happens when the bucket overflows

Tools handle overflow in one of three ways, and they rarely tell you which:

• Truncation. The file is cut off. The model reads the first however-many pages and never sees the rest. Your answer is built on the front half of the document, stated as if it were the whole thing.
• Summarization. The tool condenses the file before the model reads it. Detail is lost: exact numbers, specific clauses, the one sentence you actually needed.
• Retrieval. The tool stores the whole file, then searches it for the parts that seem relevant to your question and reads only those. This is the smartest option, and it has its own failure modes. Lesson 4 is entirely about it.

None of these is wrong. What is wrong is not knowing which one happened. An AI answer about page 80 of a document that was cut off at page 60 looks exactly like an AI answer about page 80 of a document it fully read.

The limits, as of August 2026

These numbers change often. Treat them as a snapshot from the official help pages on the date this lesson was updated, and check the help page for your tool when it matters.

• ChatGPT. Up to 512MB per file. Text and document files are capped at 2 million tokens each, spreadsheets at roughly 50MB, and images at 20MB. Free accounts are limited to 3 file uploads a day.
• Claude. Up to 20 files per chat. PDFs are capped at 1,000 pages, and the visual elements (charts, images) are analyzed only in PDFs of 100 pages or fewer. Files added to a Project are capped at 30MB each.
• Gemini. Up to 10 files per prompt. Videos up to 2GB, every other file type up to 100MB. The larger 1 million token reading window is tied to the paid Google AI plans.

Notice what those numbers do not tell you: how much of a big file the model will actually hold in view once your conversation is also in the bucket. That is gate three, and no help page gives you a number for it because it depends on the chat you are in.

A rough rule of thumb for planning: a page of ordinary business text is about 500 words, and a word is a little more than one token, so a 100-page document is somewhere around 70,000 tokens. That fits comfortably in a million-token window on paper. In practice, treat anything over a hundred pages as a candidate for the chunking methods below, because the tool in front of you may be using far less than the model's full window.

Signs you hit a limit

Because gate three fails silently, you learn to read the symptoms instead of waiting for an error. Four tells that a file was not fully read:

• Answers get vaguer the deeper you ask. Questions about the opening pages come back specific and quoted. Questions about the closing pages come back general, or restate the introduction.
• It cannot find something you know is there. You ask about the flood endorsement and it says the document does not mention one. Before you believe it, ask which page it read last.
• The summary is front-heavy. Five bullets on the first section, one bullet covering the last four sections. That shape means the tail was summarized or dropped.
• Earlier context vanishes. Twenty messages into a chat about the file, the AI asks you what document you are referring to. The conversation grew past the window and the file scrolled out of view. Nothing is wrong with the tool. Start a new chat and re-attach.

The habit that catches all four is the one from lesson 1: ask the AI to quote something from the end of the file before you ask it anything about the middle.

Chunking a long file

Chunking just means splitting a long document into pieces and working through them one at a time. It is the old-fashioned fix, and it still beats every clever feature when accuracy matters. Three ways to do it, from easiest to most thorough:

• Split by section. A 140-page policy is really ten documents stapled together. Save each section as its own PDF, or copy and paste one section at a time, and ask your question of each.
• Extract, then ask. First pass: "List every clause in this file that mentions flood, water, or storm, with the page number, quoted exactly." Second pass, in a fresh chat with only that list: "Given these clauses, is flood damage covered?" The second chat is small, so nothing overflows.
• Rolling summary. For a long transcript or a year of notes: feed it in 20-page pieces, and after each piece ask for an updated running summary of what matters. Carry that summary into the next piece. It is slower, and it is how you read 500 pages with a tool that holds 100.

One rule ties all three together: start a new chat for each chunk when the chunks are big. A conversation that already holds five chunks has no room for the sixth. The context window fills up with the whole history, not just the latest file.

Try this yourself

Find the longest PDF you own. Upload it and paste this before you ask anything else:

This document is long. Before we start:
1. Tell me the total page count you can see, and quote the last full sentence of the final page you can read.
2. If you could not read the entire file, say so plainly and tell me where your view ends.
3. Then give me a one-line summary of each major section with its starting page.
I will ask my real questions after you confirm.

If the last sentence it quotes is really the last sentence of your document, you are inside the bucket and can ask freely. If it quotes something from the middle, you just found the cut line. Split the file there and continue in a new chat.