Build a ChatGPT knowledge base from your existing files

September 5, 2026 · Ask Your Docs

A practical ChatGPT knowledge base starts with a Project containing a small set of authoritative files, clear instructions and questions you can verify. Add your reference material, ask for evidence from it, and test an update before inviting colleagues to depend on the answers. A folder full of uploads is only the starting point.

This guide uses native ChatGPT Projects for the first setup, then explains when a connected document library is useful. It includes downloadable fictional sources and an acceptance checklist. The steps are based on current documentation as of September 2026; the checklist is an evaluation method, not a claim that a particular account passed it.

Choose one job for the knowledge base

Start with a recurring question your team already answers from documents: what needs approval before a campaign launches, which product claims are supported, or how a process handles an exception. Write that question before collecting files. It tells you which sources belong in the first version and which material would only add noise.

For the example here, the job is campaign approval. A correct answer needs both the current approval status and the rules that govern it. A separate spreadsheet exercise tests numeric questions, but it is not evidence about the campaign budget. Keeping those purposes explicit prevents a fluent answer from combining unrelated facts.

Decide who owns the sources and who can approve changes. Record a filename, version, owner and review date for each important document. If two documents disagree, resolve that conflict or tell the assistant to flag it; do not leave it to invent which policy is authoritative.

Create a ChatGPT Project and add sources

OpenAI’s Projects documentation describes a workspace for related chats, reference files and instructions. Projects are available across free and paid plans, while limits and workspace controls vary. Use the options shown in your account rather than assuming another person’s subscription has the same settings.

1. Create a new project with a descriptive name such as “Campaign approval practice.”
2. Add the reference files to the project sources.
3. Open the sources list and confirm the expected filenames.
4. Add project instructions describing how answers should use evidence.
5. Start a conversation inside that project and run a known-answer question.

Begin with the campaign approval Word document. Use the separate assistant instructions for setup. Keep the review checklist outside the uploaded source set: it contains expected answers. The sales workbook adds a separate spreadsheet exercise; request source rows, not just the stored check values.

Write instructions that make answers reviewable

Answer from the supplied source files. Name the file and heading or table supporting each material claim. Separate source facts from calculations and inferences. Check exceptions before saying an action is approved. If information is missing or conflicting, say what is missing or conflicting. Do not invent a business rule.

Instructions define the requested behavior. They do not guarantee the assistant will follow it perfectly, so pair them with checks. Avoid broad requests such as “know everything about our business.” Specify what evidence looks like for the job: a policy heading, an approval-table row or the records used in a calculation.

Keep operational directions in project instructions and business facts in source files. If a source document contains a quoted prompt or an instruction addressed to an AI, it should be treated as document content, not authority to change your request. Review any output that unexpectedly changes the task or asks for unrelated access.

Run a fact, exception and missing-information test

Open the original after each response. Verify that the cited passage exists and supports the conclusion. A citation to the budget cannot support a cancellation fee. A status marked approved cannot settle a new exception without considering the additional rule. Record the answer and the actual source, including any failure.

The Word guide walks through these checks. For the numeric exercise, use the Excel guide to inspect filters and calculations. Reading a value and recalculating a workbook are different acceptance criteria.

Test what happens when a document changes

Create a clearly named version 1.1 copy of the fictional policy and change the launch date to October 19, 2026. Replace the old source in the test project, or remove it before adding the new copy. Start a new project conversation and ask for the launch date and source version. The expected answer should now cite the changed document.

This exercise tests the maintenance process, not just the initial upload. If both versions remain, ask which one was used and why. An old conversation can contain an earlier answer, so repeating that answer is not evidence that the replacement document was retrieved. Keep the current-source list separate from the history of decisions.

For real work, assign the update to a document owner and rerun the affected questions after each material change. If a policy has been withdrawn, verify that users know what replaces it. A knowledge base becomes unreliable when availability outlives authority.

Decide between a native project and a connected library

A native project is a reasonable fit when your work stays in ChatGPT and its source-management options meet your needs. A connected library becomes relevant when maintaining the same source collection across assistants is a requirement. Compare the setup and update work using the same documents, not a list of generic features.

For Ask Your Docs, upload and organize the sources, then follow Connect your AI. The current public setup includes an OAuth sign-in to connect your library. If the add-app or developer option is unavailable, check plan and workspace permissions; your administrator may control it. A connected label alone does not prove a useful answer—run the source checks again through the connection.

Also check file behavior. Ask Your Docs retrieves extracted passages; spreadsheet formulas and complete visual layouts are not implied by upload support. For PDF-specific checks, use the PDF guide. OpenAI’s upload documentation describes separate native file limits and processing rules.

Share only after the sources and access are clear

Before sharing, inspect the project’s actual sharing settings and who can view or change reference material. Ask a colleague with the intended access to run a harmless known-answer question. Treat a question answered by your owner account as proof only for that account, not for the entire team.

Keep a short record of the source versions, setup route and date of the check. Repeat it when permissions or documents change. For a parallel workflow in another assistant, use the Claude knowledge base guide with the same examples. Similar answers are useful only when each assistant can point back to the correct evidence.

Frequently asked questions