Blog
Notes on building AI that cites its sources.
Retrieval, agents, evaluation and the unglamorous engineering that decides whether any of it works.
Chunking decides your RAG quality more than your model does
Teams reach for a bigger model when answers go wrong. Nine times out of ten the retrieval never surfaced the right passage in the first place.
An answer without a citation is a guess with good grammar
Citations are not a nice-to-have on top of a RAG product. They are the only mechanism a reader has to decide whether to trust what they just read.
When an agent beats a workflow — and when it does not
Agents are the right answer when the steps depend on what the previous step found. For everything else, a fixed pipeline is cheaper, faster and easier to debug.
Permission-aware retrieval, or how not to leak the salary sheet
The fastest way to lose trust in an internal assistant is for it to answer one question it should not have been able to answer.
Evaluating RAG when you have no labelled dataset
You do not need thousands of labelled pairs to know whether a change helped. You need thirty real questions and the discipline to change one thing at a time.
What breaks retrieval quality on Vietnamese documents
Diacritics, word segmentation and mixed-language sources cause most of the recall loss teams see on Vietnamese corpora — and all three are fixable at ingestion.
Showing all 6 articles.