4 months of RAG-LCC: query-action splitting and table retrieval lessons

HarinezumIgel · reddit · 2026-11-21

A retiree building RAG-LCC shares 4 months of lessons: an orchestrator for RAG flows, static + LLM-judge confidence scoring, a TableChunker that got tables into chunks but not reliably into answers, keeping raw unrendered answers for grounding when using Rich CLI rendering, batched document summarization via a pinned-document summarize: option — and the core insight that query and query action must be split, since "create a summary" breaks retrieval.

Original post →

More from coding & agent

coding & agent channel →