Personalizing AI: Make It Work Like You Do
Why AI answers feel generic, and the four levers — instructions, context, memory, and examples — that make an assistant work the way you do. With a method for checking that it actually helped.
Projects, spaces, and your own documents
General knowledge cannot know what your team decided in March A model trained on the public internet knows a great deal about your industry and nothing at all about your organisation. The fix is not a better prompt — it is grounding: putting your own material into the context so the answer is built from […]
Teaching it your voice by showing, not telling
Describing your style barely works. Showing it works immediately. Ask for “a warm but professional tone” and you get someone’s average of those words. Paste two emails you actually wrote and ask for a third in the same voice, and the difference is obvious in one attempt. This is the examples lever, and it works […]
Grounding AI in Your Own Knowledge (RAG)
Why models invent answers about your business, and how retrieval-augmented generation fixes it by giving them your actual sources.
The problem: confident answers about things it never saw
A language model is not a database A model predicts likely text. That is a wonderful property for drafting and terrible for facts it was never trained on. Ask about your internal refund policy and it will produce something that sounds exactly like a refund policy — fluent, plausible, and invented. Three limits cause this: […]
Diagnosing a bad answer: retrieval or generation?
Two failures that look identical to the user A RAG system gives a wrong answer. Almost everyone’s first move is to edit the prompt. That fixes about half the cases and wastes time on the other half, because there are two entirely separate failures wearing the same face. Retrieval failure — the right passage was […]
Keeping the corpus honest
A grounded system is only as good as what it is grounded in RAG answers faithfully from your documents. If your documents contain three versions of the expenses policy, two of them obsolete, the system will answer faithfully from the wrong one — and it will cite its source while doing it, which makes the […]
Measuring a RAG system honestly
“It seems good in the demo” is where most RAG projects stop Grounded systems demo beautifully. Someone asks three questions they already know the answers to, the citations look convincing, and the project is declared a success. Then it meets real questions from people who cannot check the answers. Because RAG has two components, it […]
When not to build RAG
The most expensive RAG system is the one that should have been a search box RAG is genuinely the right answer for a large body of documents and open-ended questions. It is also proposed constantly for problems it does not fit, at considerable cost. Four cases where something simpler wins The corpus is small and […]