Notes
RAG is the future of answers that have to be yours.
A general model will keep getting better at language. It will not absorb the shared drive, the policy edited on Thursday, or the note from this morning’s visit. Retrieval-augmented generation is how those files get into the answer, on the day someone asks.
The position is plain. Retrieve the current, permitted passage. Show the page. Let a person reject the sentence. A fluent answer from the wrong document is worse than a short one that admits the file does not say.
These are short notes, written so the argument is public while the longer essays are still being drafted. They will be replaced by the full pieces. The stance will not soften in the meantime.
- Retrieval · 10 November 2026 RAG is how the answer stays yours A model can sound sure about a policy it has never seen. Retrieval is what puts your document in the answer.
- Retrieval · 10 November 2026 The policy is in the shared drive Training data stops on the day the model was built. Your procedure changed on Thursday.
- Retrieval · 10 November 2026 Retrieval fails quietly The dangerous miss is a fluent answer written from the wrong page.
- Retrieval · 10 November 2026 What to retrieve, and what to leave out More chunks are not more truth. A note from 2019 can crowd out the plan from last week.