LEARNING PATH
RAG knowledge engineering
Build maintainable RAG systems through chunking, retrieval, reranking, evaluation, and access filtering.
RAG knowledge engineering
Build maintainable RAG systems through chunking, retrieval, reranking, evaluation, and access filtering.
Start with a traceable document pipeline and define content, version, permission, and deletion semantics.
Open collection →TUTORIAL_CHAPTEREmbeddings and vector indexes: reproducible retrieval from experiment to productionChoose an embedding model, distance metric, and batching strategy while keeping the index rebuildable.
Open collection →TUTORIAL_CHAPTERHybrid retrieval and reranking: improve candidate relevanceCombine keyword, vector, and filter signals while keeping relevance scoring separate from permission boundaries.
Open collection →TUTORIAL_CHAPTERRAG evaluation and citations: separate retrieval from generationUse traceable question sets to measure recall coverage, citation accuracy, faithful answers, and justified refusal.
Open collection →TUTORIAL_CHAPTERRAG permission filtering and knowledge freshnessMake tenant permissions, document status, and freshness first-class fields in retrieval so unauthorized or stale evidence never reaches the model.
Open collection →TUTORIAL_CHAPTERRAG production pipeline: updates, cache, and rollbackVersion the ingestion and index pipeline so updates are observable, caches stay isolated, and an incomplete index cannot become current.
Open collection →casesCase study: a local SQLite policy-search RAGAssemble a reproducible RAG with SQLite FTS and deterministic context, without an external service.
Open collection →casesCase study: blue-green RAG index release and rollbackBuild and sample-check a new index, then switch safely with a version pointer.
Open collection →casesCase study: replaying RAG retrieval regression with a fixed question setUse source ids, permission tags, and expected evidence to separate retrieval regressions from generation problems.
Open collection →TUTORIAL_CHAPTERRAG 查询改写与拒答:先保证证据,再优化表达 (original language)用受限查询改写和证据阈值处理口语问题、无结果和不确定回答。
Open collection →