Dissecting Agentic RAG: A Component Ablation for Multi-Hop QA with a Local 7B Model

The paper presents an ablation study on the Agentic retrieval-augmented generation (RAG) system using a local 7 billion parameter model (Qwen2.5-7B-Instruct) for multi-hop question answering. The study evaluates a full agentic RAG pipeline against a single-pass dense-retrieval baseline, achieving significant improvements (EM=53.2%, F1=61.6%) compared to the baseline (EM=43.1%, F1=54.0%). Key findings indicate that fixed hybrid retrieval methods outperform adaptive routing, and that two retrieval iterations effectively capture most gains, suggesting that simpler, fixed strategies can be more effective than complex adaptive approaches in resource-constrained environments.