Eight-part design guide
Answer from your data
Better answers do not start with a smarter model. They start by putting the right evidence in front of the model. Read this as a sequence: prepare the material, find it, then check the answer.
0StartHow to make a system answer from your dataThe library metaphor, the words you need, and five conclusions to read before choosing a method for answering from your own material.1Failure and questionsWhy retrieval-augmented generation fails so easilyWhy a one-day RAG demo breaks in real work, how context gets lost, and the eight questions to answer before you pick an architecture.2PrepareAnswer quality is decided while you prepare the materialParsing, metadata, chunking by document type, parent-child and late chunking, indexing choices you cannot undo cheaply, and Korean-specific traps.3Find and answerFind the evidence, then answer from itHow contextual retrieval, hybrid fusion, and reranking cut different failures, plus the minimum rules for turning evidence into an answer.4Graph and wikiMap the relationships, or compile the knowledge firstWhen graph retrieval wins and loses, why the variants do not cost the same, and what a compiled wiki can and cannot hold.5Other pathsPut it all in, look it up live, or match the tool to the dataWhen long context beats search, where it fades, how agents and MCP read the source, and why tables, scans, and tone need different tools.6DesignDesign a hybrid system and choose where to startFour layers of knowledge, a router that sends each question down a different path, a small-to-large stack, and starting points by scale.7Measure and operateMeasure it, run it, and avoid the usual mistakesHow to separate retrieval failures from answer failures, a staged build, security and deletion, a design checklist, ten anti-patterns, and sources.