Particle.news
Download on the App Store
32 ARTICLES
·
6d ago

Late‑July Papers Push RAG Toward Plan‑First Retrieval, Structured Evidence and Layered Defenses

A cluster of arXiv papers and practitioner reports show researchers replacing broad ingestion with agentic planning, compact schema or hypergraph evidence, and multi-stage defenses to raise grounding, cut costs, and blunt poisoning attacks.

Image
Image
Image
Image

Overview

  • A wave of papers published July 28–30 documents a clear shift from single‑pass retrieval to agentic, plan‑before‑retrieve systems that pick what to fetch and where to search to improve grounding and traceability.
  • Structured evidence methods such as schema‑constrained causal graphs and hypergraphs report large efficiency gains, with one study claiming 3–20× fewer graph nodes and 8–135× fewer build‑time LLM calls than heavy graph baselines.
  • Security work presents layered, pipeline defenses that sharply reduce corpus‑poisoning success rates, for example a three‑ring TriShieldRAG configuration that the authors report cutting attack success from about 91% to about 13%.
  • Applied and production‑oriented projects show the pattern is practical: an APS facility deployment (APS‑RAG) raised strict vital‑nugget recall from 63.8% to 70.3% and OPG papers describe progressive, cost‑aware escalation designs for regulatory corpora.
  • Most claims come from arXiv preprints and blogs so independent replication, adversary‑adaptive testing, and production validation are still needed before the reported accuracy and efficiency gains are accepted widely.