How to Train ChatGPT on Your Own Data: Projects, RAG or Fine-Tuning?
Usama · November 6, 2026
You usually cannot "train" ChatGPT on your own data in the strict sense, and most businesses do not need to. When people ask for this, they mean one of three things: give ChatGPT your documents as reference material (a Project, or a Custom GPT where one already exists), connect it to your files through a retrieval system known as RAG, or fine-tune a model on example conversations. Start with the simplest one that passes a test on your real questions.
What does "train ChatGPT on your own data" actually mean?
A model like the one behind ChatGPT is trained once by its maker on a huge public dataset. Your price list, staff handbook and customer emails are not in there. So the real question is how to get your information in front of the model at the moment it answers.
OpenAI's own optimisation guide separates two problems. If the model lacks knowledge, you give it context to look up, like handing a student the textbook during the exam. If the model behaves inconsistently, in the wrong tone or format, you show it many examples so the habit is learned, like months of classes. Only the second is "training" in the technical sense. Nearly every small-business request is the first kind.
That is good news. Lookup is cheaper, easier to correct (edit the document, and the answer changes) and can point to where an answer came from.
The options side by side
| Option | What it really does | Privacy | Main cost drivers | Upkeep | Can it cite sources? |
|---|---|---|---|---|---|
| ChatGPT Project with files | Keeps files and instructions beside your chats; the model reads them while answering | Depends on your plan and data setting (see below) | Your ChatGPT subscription | You replace files by hand | Only if you ask it to in the instructions; no guarantee |
| Custom GPT with knowledge files | A shareable ChatGPT assistant with instructions and up to 20 files | Same plan rules; sharing settings matter | Subscription of each person who uses it | Hand-maintained, and the product is being retired | Only if instructed; weaker than a built system |
| RAG over your documents | Searches your own document store for the relevant passages, then asks the model to answer from them | You choose the hosting and who sees which documents; API data is not used for training by default | Build effort, model usage per question, hosting, re-indexing | Re-index when documents change; review failed questions | Yes, by design: it can show the passage and file |
| Fine-tuning | Adjusts a model's behaviour using hundreds of example inputs and ideal outputs | Your examples are uploaded to the provider | Preparing quality examples, training runs, then the provider's usage pricing for the tuned model | Re-train when the task or base model changes | No: knowledge is absorbed, not looked up |
Last verified November 2026 against OpenAI's help centre and developer documentation. Plans, limits and product names change often, so check the vendor pages linked at the end before you buy anything.
Which option should you try first?
Work up a ladder and stop as soon as the answers are good enough.
- Paste or attach the document in a normal chat. If you only need answers from one contract or one manual once, nothing else is required.
- Create a Project if the same files will be used repeatedly by you or a small team.
- Move to RAG when the document set is large, changes often, needs per-person access rules, or must answer customers on your website or WhatsApp.
- Consider fine-tuning only when the problem is style or format rather than missing facts, and a well-written prompt with examples has already failed. Our RAG vs fine-tuning guide covers that decision in detail.
OpenAI's guide says much the same: squeeze everything you can out of basic methods before adding retrieval or fine-tuning, and let the accuracy you need be the target, not the sophistication of the tool.
How do you set up a Project that answers from your files?
This takes about an afternoon, and the thinking matters more than the clicking.
- Write the questions first. List 20 to 30 real questions a customer or colleague has asked, with the correct answer beside each. This is your test set; you will reuse it every time you change anything.
- Choose files that actually contain the answers. Current policies, price lists, FAQs, process documents. Leave out old drafts, because the model cannot tell which version is current.
- Make the files easy to read. OpenAI's guidance for GPT knowledge is to prefer clear, text-forward files, since complex layouts can make content harder to use. One topic per file, plain headings, tables as real tables rather than screenshots. Its upload FAQ also notes that images embedded inside project files may not be picked up automatically.
- Write instructions that limit the model. For example: answer only from the project files; if the answer is not there, say so and suggest who to ask; name the file you used.
- Run the test set. Mark each answer right, partly right or wrong. Fix wrong answers by fixing the source file or the instruction, then run all questions again.
- Check the data setting before sharing anything sensitive. See the next section.
What are the file limits?
Per OpenAI's file upload FAQ, project file limits depend on the plan: up to 5 files on Free, up to 25 on Go and Plus, and up to 40 on Edu, Pro, Business and Enterprise. Documents can be up to 512 MB each, text and document files are capped at 2 million tokens per file, and a Custom GPT can hold up to 20 knowledge files. A file count that low is the usual reason people outgrow this approach, because a business with 300 pages of policies has to merge files and loses organisation.
Who can see what you upload?
This is the question to settle before the first upload. OpenAI's data-use page states that for its services for individuals (Free, Plus and Pro) it may use your content to train its models unless you turn off the "Improve the model for everyone" setting under Data controls or use its privacy portal. For ChatGPT Business, Enterprise and Edu, and for the API, it says it does not use inputs or outputs to improve models by default. It also notes that giving thumbs-up or thumbs-down feedback on an individual-plan response can allow that whole conversation to be used.
That means a personal account is a poor home for anything confidential, whatever the toggle says, because one careless click on a feedback button changes the picture. For a wider view of the trade-offs, see our piece on AI, personalisation and compliance.
Last verified November 2026. Re-read the data-use page before relying on it; these policies are rewritten regularly.
What should you never upload?
As general information rather than legal advice, keep these out of any general-purpose AI tool unless a lawyer or your data protection lead has approved the setup:
- Customer or patient records, especially health, financial or children's data.
- Payment card numbers, bank details, passwords, API keys and access tokens.
- Contracts with confidentiality clauses that bar sharing with a third party.
- Anything covered by legal privilege.
- HR files: salaries, disciplinary notes, medical leave.
- Other people's copyrighted material that your licence does not let you copy into another service.
- Unreleased financial results or other information your own regulator treats as restricted.
If the assistant does need personal data to do its job, that is a signal to build it properly with access rules, logging and a data processing agreement, not to upload a spreadsheet and hope.
Custom GPTs: check their status before you build one
A Custom GPT is a shareable version of ChatGPT with its own instructions and knowledge files. It used to be the obvious answer to this question, but OpenAI's help centre now says it plans to retire custom GPTs and move creators to plugins (see its retirement and migration FAQ). As of November 2026 its page states that new GPT creation and publishing are not available on personal accounts (Free, Go, Plus, Pro), that standard retirement is scheduled for 11 December 2026, and that GPT knowledge files are copied into a plugin's reference files during migration. Custom actions do not transfer automatically.
The practical advice: do not start a new business process on a Custom GPT. If you already rely on one, read OpenAI's retirement FAQ and plan the migration. This is also a good example of why we say to keep your source documents in your own folders rather than only inside a vendor's tool.
Three small-business examples (illustrative)
These are made-up scenarios to show the reasoning, not client results.
A letting agency with a tenant guide
A small agency has one 30-page tenant guide, a fees sheet and six template emails. Staff keep asking each other the same questions. A shared Project with those files and a "say when you do not know" instruction is probably enough. Nobody outside the team uses it, nothing changes daily, and the files are well under the plan limits.
An online shop with a changing catalogue
A store has a returns policy, a sizing guide and several hundred products whose stock and prices change daily. A Project would go stale within a week and could not sit on the shop's website. This one needs retrieval connected to the live catalogue and policy pages, with an answer that links to the product or policy it used. That is a RAG build, and the line between this and a plain chatbot is explained in our comparison of Custom GPTs, RAG chatbots and AI agents.
A consultancy that wants proposals in its house style
A consultancy wants first drafts to sound like its senior partner. The facts for each proposal change every time, so lookup is not the issue; voice is. Start with a prompt that includes three good proposals as examples. Only if that fails across a measured batch of 30 drafts is a fine-tune worth considering, and the team should check which providers still offer one, because OpenAI has closed its self-serve platform to new customers.
When file uploads stop being enough
These are the signals that a purpose-built assistant is worth the money:
- You have more documents than the file cap, or the documents change weekly.
- Different people must see different documents, such as sales versus finance.
- Customers, not just staff, will ask the questions, on your site, in WhatsApp or by email.
- You need an audit trail: which passage produced which answer, and who asked.
- You want the assistant to do something afterwards, such as create a ticket, which makes it closer to an AI agent than a chatbot.
If three or more apply, scope a retrieval-based assistant. Our custom AI assistant service covers building one that answers from your documents with source citations, and the build can start from the test set you wrote above. If the assistant will live on your website, our AI chatbot integration work is the other half.
Not sure which option fits your documents?
Tell us what you want the assistant to answer and who will use it, and we will say whether a Project is enough.
Talk through your use case →Frequently asked questions
Can I train ChatGPT on my own data for free?
You can attach files to a Project on the Free plan, but OpenAI limits it to 5 files per project and 3 file uploads a day. That is fine for a trial with one or two documents and too tight for a real knowledge base.
Does ChatGPT learn from the files I upload?
Not in the sense of permanently absorbing them. The files are read as context when you ask a question. Separately, whether OpenAI may use your content to improve its models depends on your plan and your data-controls setting, covered above.
Is RAG better than fine-tuning for company documents?
For facts that change, yes in almost every case, because you update a document instead of re-training and the system can cite where an answer came from. Fine-tuning suits consistent tone, format or classification.
Can I train ChatGPT on my website content?
You can export key pages as text or PDF and add them to a Project. For a site that changes often, a retrieval system that re-reads the pages on a schedule is more reliable than manual re-uploads.
How do I know the answers are right?
Keep the question-and-answer test set from step one and re-run it after every change. Treat any assistant as a draft-maker for staff until it has passed that test over a few weeks, and keep a human in the loop for anything with legal or financial consequences.
Related Services