Pasting an internal PDF into ChatGPT is fast. The document leaves the perimeter. The model does not have your identity. There is no source a colleague can open. Tomorrow there is no system. It is a habit, not a setup.
What breaks when you paste
- Perimeter. The file goes to a public service. Even with “do not train on my data”, it is still a send nobody designed.
- Who sees what. Whoever has the PDF pastes it. Whoever should not see it has no vote. The model does not inherit the Drive or wiki ACL.
- Source. The answer does not link to the current procedure. The next day nobody knows which version was used.
- Operations. There is no index. No reindex. No question log. Each person repeats the copy and paste.
The same applies to Custom GPTs with documents uploaded to a personal account. It is still the file inside a product that is not your environment.
What to set up instead
A RAG or GPT of your own. The documents live where you say. The answer cites. The chat or API uses company accounts. Someone can reindex. That is in set up an internal RAG and in data and access for a private GPT.
The case does not replace every PDF in the archive. It replaces the habit in one case: “this kind of question is no longer pasted; it is asked of the system”. The document stays in the wiki or Drive. The model only sees what is indexed.
How to leave the habit
- Name the case (one team, one document type).
- Stop pasting that kind of file into public chats.
- Put the chat or API where people already work. If they have to open another forgotten site, they will go back to pasting.
- When the answer fails, you fix the source document and reindex. You do not “teach” the chat with more loose PDFs.
OpenAI can be the model. LangChain, Python and FastAPI, the wiring. The PDF stays where it is. Madrid. RAG and agents.