How to build a knowledge base in a ChatGPT Project

A project is best understood as a small reference library with a conversation attached. The files persist, the chats come and go. When you start a new chat inside the project, you get a clean context window but the same documents, which is exactly what you want when you are asking about the same body of material week after week.

The mistake is treating it as storage. A project is not a folder for everything you might need one day. It is a working set, and the quality of what you get back is tied closely to how disciplined that set is.

What belongs in project files, and what stays in a chat

The dividing line is repetition. If you will ask about a document more than a few times, it belongs in the project. If it is today's problem, attach it to the chat.

The second column matters more than people expect. Every extra document in the project is another candidate answer to every question you ask, including the questions it has nothing to do with.

The ceilings you are working inside

Five separate limits apply, and they fail in different ways.

Building a project of thirty files in one sitting spends thirty slots against the rolling rate, so a bulk import can stop halfway through for reasons that have nothing to do with the project. Storage is cumulative across everything in your account and never clears with time.

Why a smaller set answers better

OpenAI does not publish how project files are selected when you ask a question, so treat any confident explanation of the retrieval mechanism with suspicion, including this one. What is observable is the failure pattern, and it is consistent.

Crowded projects go wrong in three ways. Near duplicates are the worst: keep version 3 and version 7 of the same policy in a project and answers will come from one of them with no reliable way to tell which. Broad questions across many documents produce blended answers where attribution disappears. And dead weight, files that contributed no readable text, makes the project look better stocked than it is.

The working rule: one document per topic, one current version of each, and nothing that nobody will ask about.

Preparing files before they go in

Ten minutes of preparation is worth more than any prompt you will write afterwards.

1. Name every file for what it contains

   expenses-policy-2026.pdf rather than final_v3_FINAL.pdf. Filenames travel with the document and you will be referring to them out loud in your questions, so make them say something.

2. Merge related short documents

   The limit counts files, not words. Six short procedure notes become one operations reference, with a clear header before each original so the boundaries survive the merge.

3. Convert to plain text or markdown where layout does not matter

   Smaller, cleaner, and it removes images that were going to be discarded anyway.

4. OCR anything scanned

   Or accept that it contributes nothing.

5. Remove the old version when you add a new one

   This is the habit that keeps a project sharp over months. Superseded documents are the single most common cause of a project quietly going wrong.

Use the project instructions to describe the library

A project carries its own instructions, and this is where you say what the documents are and how they relate to each other. Something like:

These files are our client contracts currently in force, one per client, named by client. When answering, name the file the answer came from and quote the clause. If the files do not cover a question, say so rather than generalising from the others.

That does two jobs. It tells ChatGPT how to attribute answers, which is what makes a knowledge base trustworthy rather than merely convenient. And it gives a sanctioned way to say the answer is not there, which cuts down on confident invention.

Keeping it current

A knowledge base decays. Put a recurring twenty minutes in the diary, monthly is plenty for most projects, and do three things: delete anything superseded, check the file count against your plan's ceiling, and check Settings then Storage against the 25 GB cap.

Then test it. Ask three questions you already know the answers to and see whether the citations point at the right documents. If they point at the wrong file, you have a duplicate to remove.

The practical decision

Use a project when the same documents come up again and again, and keep the set deliberately small. Repeated reference material goes in the project, today's file goes in the chat, and every document gets a name that says what it is. Delete the old version when the new one arrives. That single habit does more for answer quality than anything else on this page.

If instead you want to hand a fixed reference to other people through one assistant, a custom GPT is the better shape, though its ceiling of 10 files per GPT covers the lifetime of that GPT and never refreshes. It is also only open to you inside a Business, Enterprise or Edu workspace, because GPT creation has been withdrawn from personal accounts. Projects are for knowledge bases that keep changing, and on a personal plan they are the only option of the two.