How to Add a Knowledge File to a Custom GPT (And Actually Make It Work)

What Happens When Your GPT Doesn't Know Your Credit Union?

You built a Custom GPT. You gave it instructions. You asked it to help draft a member response about your overdraft policy—and it handed you something polished, competent, and completely generic. Something that could've come from any bank in the country.

That's the gap a knowledge file closes. Upload the right document, ask the right question, and your GPT stops guessing and starts quoting your actual policy.

What Is a Knowledge File in a Custom GPT?

When you build a Custom GPT inside ChatGPT, you can upload documents directly into the GPT's configuration. Those documents become a private reference library. When someone on your team asks the GPT a question, it searches those files and pulls relevant context before it answers.

Think of it like handing a new employee your policy manual, your rate sheet, and your member handbook on day one—except the GPT actually reads all of it before responding.

A knowledge file doesn't replace the model's general intelligence. It gives the model something specific to work with—your words, your policies, your voice.

This is called Retrieval-Augmented Generation, or RAG. The GPT retrieves relevant passages from your uploaded documents, then uses those passages to shape a more accurate, more specific answer. You don't need to understand the technical plumbing. You just need to know what it does—and what trips it up.

What File Types Can You Upload to a Custom GPT?

As of 2025, ChatGPT's Custom GPT builder accepts a range of common file types. The ones most useful for credit unions include:

• PDF — your best option for policy documents, member disclosures, loan guidelines, and anything that already exists as a formatted document
• Word (.docx) — drafts, procedures, internal memos
• Plain text (.txt) — clean and simple, no formatting overhead
• Markdown (.md) — useful if your team works in tools like Notion
• PowerPoint (.pptx) — works, but text-heavy slides pull better than visual ones
• Spreadsheets (.csv, .xlsx) — usable for rate tables or product comparison data, though the GPT handles narrative prose more reliably than rows of numbers

Always verify current file-type support at platform.openai.com—this list changes as OpenAI updates the product.

Each file can be up to 512 MB, and you can upload up to 20 files per Custom GPT (verify current limits at openai.com, as these have shifted over time). In practice, clean and focused files outperform large, sprawling ones. A tightly written 10-page lending FAQ will outperform a 90-page all-hands policy dump every time.

How Do I Add a Knowledge File to a Custom GPT?

You'll need a ChatGPT Plus, Team, or Enterprise account to build Custom GPTs—verify current plan requirements and pricing at openai.com.

1. Open the GPT builder. Go to chatgpt.com and click your profile icon in the top-right corner. Select "My GPTs," then click "Create a GPT" or open an existing one to edit it.

2. Switch to the Configure tab. The builder has two tabs: Create (a conversational setup tool) and Configure (where the real controls live). Click Configure.

3. Scroll to the Knowledge section. Partway down the Configure page, you'll see a section labeled "Knowledge." Click "Upload files."

4. Select your file. Choose a PDF, Word doc, or other supported file from your computer. Give it a moment to upload and process. It will appear as a thumbnail below the upload button.

5. Enable "Code Interpreter & Data Analysis" if needed. If you're uploading spreadsheets or want the GPT to do any math with your data, check that toggle. For most policy documents and narrative text, you can leave it off.

6. Save and test immediately. Click "Update" (or "Create" for a new GPT), then open a conversation and ask a specific question that your document should answer. Don't assume it works—confirm it.

Before and After: What a Knowledge File Actually Changes

Say your credit union has a specific policy: first-time homebuyers who qualify for your Community Champions mortgage product get a 0.25% rate reduction if they complete a homebuyer education course.

Without a knowledge file, you ask your GPT: "What should I tell a first-time homebuyer about our rate discounts?"

It says something like: "Many lenders offer rate discounts for first-time homebuyers. You may want to mention any applicable programs, such as down payment assistance or rate reductions tied to education requirements."

Accurate? Sure. Useful to your loan officer? Not really.

With your lending FAQ uploaded, you ask the same question:

The GPT says: "First-time homebuyers who qualify for the Community Champions mortgage product can receive a 0.25% rate reduction by completing an approved homebuyer education course. Confirm they meet the income and property eligibility criteria first."

Same model. Same question. Entirely different usefulness.

The Mistake Most People Make

Here's what goes wrong more often than not: someone uploads a file and walks away expecting magic. They ask the GPT a question, get a mediocre answer, and conclude the knowledge file "didn't work."

Usually, the file worked fine. The question was too vague.

The GPT retrieves content based on how closely your question matches the language in the document. If your policy doc calls it a "rate adjustment" and you ask about "discounts," the retrieval may miss. If you ask something broad like "tell me about our mortgage products," the GPT has no anchor—it doesn't know which part of a 40-page document to pull from.

The fix is straightforward: ask specific questions that use the same terminology your document uses. Coach your member services team or loan ops team the same way. "What is the income limit for the Community Champions product?" will pull a cleaner answer than "what are our mortgage options?"

A second common mistake: uploading scanned PDFs that are images, not text. The GPT cannot read a picture of a document. Run scanned files through OCR software first, or copy the key content into a clean Word doc before uploading.

Why This Matters for Your Credit Union

Your policies, products, and rates are not generic. Your members deserve answers that reflect your actual terms—not a best-guess composite of what the industry typically does.

A well-configured Custom GPT with a solid knowledge file is one of the most practical AI moves your team can make right now. Your member services team gets answers grounded in your actual disclosures. Your loan ops team can check guidelines without digging through shared drives. Your marketing team can write product descriptions that match what you actually offer.

It's not complicated. It takes the right file, the right questions, and about twenty minutes of setup.

This post was drafted with AI assistance and reviewed by a human at CU 2.0. AI makes mistakes; verify any specific claim before acting on it.

What's the risk?

The biggest risk here is data exposure. When you upload a file to a Custom GPT, that content is processed and stored by OpenAI on their infrastructure. For internal training documents or general product FAQs, that's a manageable risk most credit unions can live with. For documents containing member NPI, nonpublic financial data, or sensitive regulatory correspondence—talk to your compliance team before uploading anything. Review OpenAI's data handling and retention policies for your specific plan tier (Plus, Team, and Enterprise have different terms), and consult your information security officer if you're unsure where a document falls.

The accuracy risk is lower than many executives assume, but it's real. The GPT can misquote or partially retrieve a passage, especially from long or poorly structured documents. Never let a knowledge-file-powered GPT answer a member directly without a human review step in the workflow. For internal use—loan officers checking guidelines, your board prep team summarizing policy options, marketing drafting product copy for human review—the risk profile is low and the practical value is high. Go build it. Just don't upload your SAR log in the process.

Frequently Asked Questions

What file types can I upload to a Custom GPT?

As of 2025, ChatGPT supports PDF, Word (.docx), plain text (.txt), Markdown (.md), PowerPoint (.pptx), and spreadsheets (.csv, .xlsx), among others. PDFs and Word docs tend to work best for credit union policy and procedure documents. Always verify the current list at platform.openai.com, as OpenAI updates this regularly.

How many files can I upload to a Custom GPT, and how large can they be?

OpenAI currently allows up to 20 files per Custom GPT with a 512 MB limit per file. These limits have shifted over time, so confirm current specs at openai.com before planning a large upload.

Why isn't my Custom GPT using the file I uploaded?

The most common reasons are that your question doesn't closely match the language in the document, the file is a scanned image rather than real text, or the document is too long and unfocused for the model to locate a relevant chunk. Try asking a more specific question using the exact terminology from your document.

Can I upload a scanned PDF to a Custom GPT?

Not effectively. Scanned PDFs are images, and the GPT cannot extract text from them the way it can from a text-based PDF. Run them through OCR software first, or copy the key content into a Word document before uploading.

Is it safe to upload credit union policy documents or member disclosures to a Custom GPT?

For general internal documents and product FAQs, the risk is manageable—but consult your compliance team before uploading anything containing member NPI or sensitive regulatory content. Review OpenAI's data handling policies for your specific plan tier, since Plus, Team, and Enterprise accounts have different terms.

This post was drafted with AI assistance and reviewed by a human at CU 2.0. AI makes mistakes; verify any specific claim before acting on it.