Part 7 argued that a groundedness gate has to score atomic claims and take the minimum, because a five-sentence answer with one fabrication averages out to a passing score. It then moved on, having left the hardest part unexamined: what exactly is a claim, and how do you get one out of a paragraph reliably enough that a threshold means anything? That is this post — the first that goes back to ...
Seven posts of machinery, and none of it is trustworthy until you can answer one question: how do you know it works? RAG systems are unusually hostile to intuition here, because a broken one keeps returning fluent, confident, well-cited answers. There is no crash, no 500, no red line on a dashboard. This post covers how you measure the thing, how you operate it safely, and the consolidated list...
Retrieval can succeed completely and the answer can still be wrong. The model receives five relevant chunks and asserts a sixth thing that none of them support — a plausible number, a policy that does not exist, a qualifier it invented to make the answer read better. This is a different failure from anything in the previous six posts, and it needs its own machinery, because nothing upstream can...
At some point in every RAG project, a category of question stops working and no amount of tuning fixes it. Better chunking does not help. A stronger reranker does not help. The reason is that two completely different failures both present as “the system could not answer”, and only one of them is a retrieval-quality problem at all. Part of the Production RAG series (9 parts) Production RAG -...
Text retrieval works because both sides of the comparison are text living in one embedding space. An image breaks that symmetry, and the way you restore it is one of the few RAG decisions that is genuinely hard to reverse later. It is also the decision where the cheapest-looking option is usually the expensive one, for a reason that only becomes visible when you separate one-time costs from rec...