How to work with many documents in one ChatGPT chat

Two documents in a conversation is manageable. Ten is a different problem, and it fails quietly, because the answers keep sounding fine.

What breaks is attribution. Ask a question across a pile of files and you get one confident paragraph with nothing indicating which file each sentence came from. Sometimes the sentence is stitched from three of them. Sometimes it comes from the one file that happened to match your phrasing, and the other nine were never consulted at all.

What actually happens when you attach ten files

Your documents are not all sitting open in front of the model. Long files are retrieved from: when you ask something, passages that match the question get pulled in, and the rest stays outside the conversation until something else calls for it.

That is why 2 million tokens per document is misleading if you read it as a reading capacity. It is a ceiling on what one file can contain before it is truncated, not a promise about how much of it informs any given answer.

The consequence for many documents is specific. Retrieval competes. Whichever passages best match the words you used will win, so a document that makes the same point in different vocabulary loses, and you never see that it lost.

Two further limits shape the session. Uploads are capped at 80 files every 3 hours across your account, and failed attempts count toward that, so a batch that half fails still costs you. Documents are also handled as text only, except Enterprise, meaning charts and scans inside those files contribute nothing on any plan below Enterprise.

Label every document as you attach it

Attach the files, and in the same message tell ChatGPT what each one is.

Attaching three files. Label them SUPPLIER-A (their 2025 contract), SUPPLIER-B (their 2025 contract), POLICY (our internal procurement rules). Do not analyse anything yet. Confirm you have all three and list them back with the labels.

Short labels, distinct from each other, in caps so they stand out in later answers. Filenames make terrible labels, because half of them are some variant of Final_v3_signed(2).pdf and the model will refer to them in ways you cannot follow.

Order them deliberately

Nothing OpenAI publishes describes upload order as significant. In practice it behaves as though it is, consistently enough to plan around.

Two rules cover most cases. Put reference material first, because whatever is read early tends to frame how everything after it is interpreted: the rubric, the style guide, the policy, the brief. Then put the material you want judged against it closest to the question you are about to ask.

Attach in small groups rather than one enormous message. Three files with a short note about what they are, then the next three. Each group gets a moment of the model's attention, which a single ten file dump does not get.

Refer to documents by name in every question

This is the habit that does the most work, and it costs nothing.

Never say "the second one" or "the other document". Say SUPPLIER-B. Then ask for the answer to be structured per document rather than as one merged view.

For each of SUPPLIER-A and SUPPLIER-B, give the termination notice period and the clause it comes from. One line each. Write "not found" if it is not in that document.

The tell for a merged answer is simple: no labels in the output. If you asked for per document answers and got a paragraph, the sources were blended and the answer is not checkable. Ask again, naming the files.

Keep an index in the chat

After the last upload, ask for a numbered index of every document with a one line description of what it contains. That list becomes the thing you point at for the rest of the session.

It also survives length. Long conversations degrade, and the early messages are the first to lose their grip, so pasting the index back in after a long stretch is a cheap way to re-anchor which documents exist and what they are called.

When to split into separate conversations

Three signals say stop adding files and open a new chat.

Answers stop naming documents even when you ask them to. Corrections do not stick, so a mistake you fixed ten messages ago comes back. Or the documents have quietly separated into two different questions that happen to share a folder.

The useful rule is one chat per question, not one chat per project. If you are asking about contracts, the marketing deck does not need to be there, and its presence only gives retrieval more ways to answer the wrong question.

For a document set you return to repeatedly, use a Project instead. Files live at the project level, 25 files per project and 40 files per project on the higher tiers, and every chat inside can reach them. You get a clean conversation each time without re-uploading the shelf.

What this comes down to

Many documents in one chat works when you take responsibility for identity and order. Label each file as it goes in, put reference material first, name the files in every question, and demand per document answers so you can see when one has been ignored.

When the answers stop naming names, that is the signal. Split the work, summarise each document into the same structure, and compare the summaries rather than the originals.