It doesn't know my company
A model brilliant about the world and blank about my company — and why bolting on retrieval is plumbing, not the real problem: trust in the answer
The model can explain a Kafka rebalance to me in three paragraphs. It cannot tell me the name of our staging environment.
That gap is the whole story of this autumn. I sat down last week to get help with an internal tool — our own conventions, our own deploy script, the way we name things at the studio after twenty-odd years of habit. The answer came back fluent, confident, and wrong in the specific way that only matters once you ship it. It invented a config flag. It assumed a folder structure we abandoned years ago. Brilliant about the world, blank about us.
Everyone has landed on the same fix at once. Retrieval. You take your docs, your tickets, your codebase, slice them into chunks, and feed the relevant ones in alongside the question. RAG, the acronym of the season. I've wired a version of it together and it helps — the model stops guessing and starts quoting. For a week it feels like magic.
Then you read the answer closely. It pulled the right paragraph but reasoned past it. It cited a doc that was true in March. It blended two sources into something neither of them said.
So the problem was never really retrieval. Stuffing context into the window is plumbing. The hard part is the thing underneath: can I trust what comes back, and can I show why I should. Where did this come from. Is it current. Who signed off.
I keep circling this. A model that knows everything except what's mine isn't knowledge — it's a very persuasive stranger. The leverage isn't in making it smarter. It's in giving it ground it can stand on, and giving me a reason to believe it's standing there.