https://alexkorol.github.io/lab-notes/Measured notes on RAG, agents, evaluation, and production LLM systems.en-ushttps://alexkorol.github.io/lab-notes/hybrid-retrieval-wrong-turn/https://alexkorol.github.io/lab-notes/hybrid-retrieval-wrong-turn/Fri, 10 Jul 2026 12:00:00 GMTI added BM25 to a dense RAG retriever, watched recall@1 regress, and used a cross-encoder to turn a wider candidate pool into a measurable win.https://alexkorol.github.io/lab-notes/judge-uncertainty/https://alexkorol.github.io/lab-notes/judge-uncertainty/Thu, 09 Jul 2026 12:00:00 GMTA single judge call hid variance. Three runs per rubric cost more, but exposed disagreement and created a concrete human-review gate.https://alexkorol.github.io/lab-notes/portfolio-rag-budget/https://alexkorol.github.io/lab-notes/portfolio-rag-budget/Wed, 08 Jul 2026 12:00:00 GMTThe model call is optional; retrieval, citations, refusal checks, and a useful degraded answer continue when quota or configuration fails.