Overview
- Three arXiv preprints on Retrieval‑Augmented Generation introduced SRAG, M‑RAG and HDRR on Tuesday, outlining three ways to move beyond simple chunk search.
- SRAG adds topics, sentiment, query and chunk types, knowledge graph triples and semantic tags to both queries and passages, and the authors report a 30% rise in answer scores judged by GPT‑5.
- M‑RAG replaces chunks with key‑value markers that use a light key for matching and a rich value for generation, and it outperformed chunked baselines on LongBench tasks under tight token budgets.
- HDRR first routes a question to likely source documents and then runs scoped chunk retrieval, and on the FinDER finance benchmark it recorded a 7.54 average score with a 6.4% failure rate and a 20.1% perfect‑answer rate.
- The papers target known flaws in chunked RAG such as fragmented facts and retrieval noise, yet the results are self‑reported preprint claims that rely on LLM‑as‑judge or select benchmarks and need independent validation.