Build an evaluated RAG assistant
A grounded assistant with measured retrieval, judged answers and live monitoring.
24 lessons, about 4.2 h. Anything you already know, you can prove and skip.
0 of 24 lessonsquests defended, 1 open now
- The request path and where time goesOpens later
- Tables, keys and modellingOpens later
- Observability: logs, metrics, tracesOpens later
- Open
- Calling a model: messages, tokens and costOpens later
- Prompts that ground answersOpens later
- Structured output and validationOpens later
- The RAG pipelineOpens later
- Lexical search and BM25Opens later
- Embeddings and similarityOpens later
- Vector search, exact and approximateOpens later
- Chunking long documentsOpens later
- Ground truth for searchOpens later
- Hit rate, MRR and other metricsOpens later
- Hybrid search and rank fusionOpens later
- Query rewriting, thresholds and the honest no-answerOpens later
- Offline and online evaluationOpens later
- LLM as a judgeOpens later
- Logging every LLM callOpens later
- Feedback and online judgesOpens later
- Dashboards and traces for LLM appsOpens later
- Testing strategyOpens later
- Configuration, secrets and containersOpens later
- From notebook to serviceOpens later