Overview

• RA-RAG introduces source reliability estimation, retrieves from top‑κ trustworthy sources, and aggregates evidence via weighted majority voting, with code released by the authors.
• Health‑domain experiments find adversarial documents sharply reduce alignment, though helpful evidence in the pool can preserve robustness, underscoring the need for retrieval safeguards.
• A proposed monitoring framework builds parallel deterministic and LLM‑generated knowledge graphs and measures structural deviations in real time to flag semantic anomalies and hallucinations.
• A survey of GraphRAG details graph‑structured knowledge, graph‑based retrieval with multihop reasoning, and structure‑aware integration for domain‑customized LLMs, with resources compiled online.
• A developer’s Java RAG evaluation reports severe retrieval and ranking failures on a ‘dishwasher’ query (precision 0.2, recall 0.011, MRR/nDCG/hit rate 0.0), with recommended mitigations including hybrid retrieval, reranking, and domain‑adapted embeddings.