Overview
- A new deep-dive showed a RAG pipeline returned a confident, fully supported answer from the wrong source after weak retrieval led the system to fall back to web search.
- The miss stemmed from document chunking that split the answer across pages, and a parent‑child retrieval pattern fixed it by using small chunks to find hits and passing their larger parent sections to write the answer.
- The author urged teams to log every step in the pipeline and to treat low retrieval relevance as a hard failure even when overall confidence looks high.
- A Microsoft case study reported stronger consistency, relevance, and traceability after inserting a knowledge‑graph constraint layer that filters the search space before RAG runs.
- Taken together, the reports signal a move toward hybrid, relationship‑aware systems for structured domains where rules, hierarchies, and explainable mappings matter.