Add a Knowledge Base

Status as of November 5, 2026: GPTs are in their final weeks. OpenAI is retiring custom GPTs on December 11, 2026 (February 11, 2027, for certain Enterprise workspaces with an exemption). Personal, Free, Go, Plus, and Pro accounts can no longer create or publish new GPTs; in Business, Enterprise, and Edu workspaces, creation depends on settings, and it ceased on October 26, 2026, in affected Enterprise workspaces. Existing GPTs remain usable and editable if your plan allows, until retirement; your conversations with them will remain accessible afterward. The migration (Migrate to plugin, in My GPTs) converts instructions into a skill, the knowledge base into reference files, and connected applications into apps; custom actions are not transferred. This lesson explains how GPTs work to help you maintain them, migrate them, and adapt their principles to a plugin (Lessons 15–17).

Give your GPT specialized memory

The system instructions you wrote in the previous lesson define your GPT’s behavior: who it is, how it responds, and what it refuses. But they don’t give it any new knowledge. Your collective bargaining agreement, product catalog, internal procedures, regulatory database: ChatGPT knows nothing about them, and it has no way to guess. That is precisely the role of the knowledge base, the Knowledge section in the configuration view. By uploading your documents there, you turn a brilliant generalist into an expert in your field—an expert that cites your texts instead of improvising.

How Knowledge works

The principle seems simple: you upload files, and your GPT can consult them to answer questions. Accepted formats cover most professional use cases: documents (PDF, DOCX, TXT), spreadsheets, presentations, structured data, and images. The limit is 20 files per GPT, each subject to standard import limits (up to 512 MB per document, and 2 million text tokens per file)—more than enough to cover a department’s or product’s documentation.

What happens behind the scenes deserves your attention, because it explains all the preparation tips that follow. Your GPT doesn’t memorize your files: it uses a mechanism called Retrieval-Augmented Generation (RAG). Your documents are first split into segments. When a user asks a question, the system retrieves the segments most relevant to that question, and then ChatGPT generates its response based on them. In other words, the GPT never “reads” your entire document: it only sees the excerpts the search returns. If your document is poorly structured, the search will return bad excerpts, and the response will be poor—no matter how good your instructions are. The quality of your documents directly determines the quality of the responses.

Prepare your documents

This RAG mechanism dictates your preparation method. First reflex: structure. A document full of clear headings and subheadings splits into coherent segments; a 500-page block without hierarchy splits at random. Also prefer one document per theme rather than a monolithic file: if your HR GPT needs to answer questions about leave and expense reports, two separate files will yield more precise search results than an “all-in-one guide.” For long documents that remain necessary, a table of contents at the start helps the system navigate.

Second reflex: write for search. Write complete sentences rather than isolated keywords—a segment that says “The waiting period between two fixed-term contracts is half the duration of the initial contract” will be found and used much better than a table cell containing “waiting period: 50%.” In the same spirit, integrate frequent questions and their answers directly into your documents: when a user asks a similar question, the match is nearly immediate. Finally, beware of complex tables, which RAG handles poorly once split: convert them to lists or structured text whenever possible.

To give you some benchmarks, here are typical documents by GPT type:

Your turn: an onboarding assistant

A clean knowledge base. Imagine you are building an “Onboarding Assistant” GPT to welcome new hires at your company. An effective knowledge base would fit into five files: the complete onboarding guide (guide-new-employee.pdf), the 30 most frequent questions from new hires (faq-hr.md), the organizational chart (org-chart.pdf), tutorials for internal tools like Slack, Jira, or GitLab (internal-tools.md), and a glossary of house vocabulary and acronyms (glossary.md). Five thematic files, each with a clear scope—exactly what RAG exploits best.

Now link the Knowledge to the instructions. Add a directive like: “When asked a question, consult your knowledge files first. If the answer is not there, state that and offer a response based on your general knowledge.” Without this directive, the GPT might answer from memory even though the correct answer was sleeping in your files.

Pitfalls to avoid

The first pitfall is uploading raw files: a 500-page PDF export without headings will yield disappointing results, and you will wrongly conclude that “Knowledge doesn’t work.” Take the time to structure before uploading. The second is the opposite of sobriety: the allowed 20 files are not a target. Five well-built documents beat twenty mediocre ones, and a catch-all file mixing topics makes search imprecise. A third, more insidious pitfall: obsolescence. Your documents are living—if your remote work policy changes, the file in the GPT must change too, otherwise your assistant will confidently answer with the old rule. Finally, a point of vigilance that isn’t technical but is critical: files uploaded to a public GPT are accessible via that GPT. Never place confidential or personal data in the Knowledge base of a GPT you plan to publish.

Official sources: Uploading files and audio to ChatGPT (OpenAI Help Center: 20 files per GPT, import limits) · GPTs in ChatGPT (Knowledge) · Custom GPT retirement and migration FAQ (files copied to plugin)

Key takeaways

• Knowledge gives your GPT access to your specific documents via a RAG system
• Response quality depends directly on the quality and structure of your files
• Prefer several thematic documents over a single large file
• Up to 20 files per GPT, each subject to import limits; upon migration, they become the plugin’s reference files
• Watch confidentiality: do not put sensitive data in a public GPT