A training-free RAG pipeline (MAIN-RAG) that inserts three cooperating LLM agents - a Predictor that drafts per-document answers, a Judge that scores each document-query-answer triplet for support, and a Final-Predictor that answers from only the passing documents - to remove noisy retrieved passages before generation. Its distinctive idea is an adaptive relevance threshold derived from the score distribution of each query's retrieved set instead of a fixed cutoff, trimming irrelevant documents while preserving recall.
MAIN-RAG (Multi-Agent Filtering Retrieval-Augmented Generation) is a training-free RAG pipeline that uses three cooperating LLM agents to remove noisy retrieved documents before generation. Agent-1 (…