I have a large repository of notes, article drafts, and commonplace book-type stuff. I experimented a year or so ago with a system using RAG to "ask myself" what I have to say about various topics. (I suppose nowadays I would use MCP instead of RAG?) I was not especially impressed by the results with the models I was able to run: long-winded responses full of slop and repetition, irrelevant information pulled in from notes that had some semantically similar ideas, and such. I'm certainly not going to feed the contents of my private notebooks to any of the AI companies.
You'd still use RAG, just use MCP to more easily connect an LLM to your RAG pipeline
To clarify: what I was doing was first querying for the documents via a standard document database query and then feeding the best matching documents to the LLM. My understanding is that with MCP I'd delegate the document query from the LLM to the tool.
As a beginner, I also haven't had much luck with embedded vector queries either. Firstly, setting it up was a major pain in the ass and I couldn't even get it to ingest anything beyond .txt files. Second, maybe it was my AI system prompt or the lack of outside search capabilities but unless i was very specific with my query the response was essentially "can't find what youre looking for"
What were you trying it in? With openwebui RAG pretty much worked out of the box.