The wiki and Drive already have the answers. The problem is finding them, knowing if they are current and citing them. Going from a folder to “questions with a source” is a pipeline. Not a prompt.
One origin per case
You pick one origin, not five. A procedures wiki, or one Drive folder, or a ticket export. Mixing everything at once produces citations to documents nobody maintains.
You need a content owner. If the wiki is dead and the knowledge lives in chats, the RAG will not revive it. First you write what you will index. Or you bound the case to what is already written.
Ingestion
- A connector or a periodic export. Not a manual zip every time someone remembers.
- Cleanup: wiki HTML, Drive comments, empty files, duplicates.
- Metadata on the way in: URL, title, date, folder, who may see it. Without a URL there is no source to open.
- Reindex when the origin changes. An index from months ago is another ghost wiki.
Python or Node for the connector. n8n or Make if the export already exists and you only need to fire the job. Docker so the job does not live in a laptop cron.
Answer with a source
“With a source” means the user sees the document and the fragment, and can open it. If the model summarises without a link, you have not left ChatGPT. Search must return identifiers, not only text.
- One answer, several sources if they apply. The model does not invent a citation.
- If there is no relevant chunk, say so. “It is not in the documents” is better than an invented procedure.
- The UI can be a chat or an API. In both, the source travels with the answer.
What not to ask for up front
That it read the whole Drive history. That it “understand the company”. That it replace the person who writes the wiki. The deliverable is: this origin, these questions, these citations, this channel. The rest comes later, if this case is used.
Sister note: set up an internal RAG. If you want me to leave it running: RAG and agents. Madrid.