Skip to content
Book Open access

Role-Specific Agentic RAG: Asymmetric Hybrid-Path Reasoning over Semi-structured Knowledge Bases

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 38 references

Abstract

Hybrid Question Answering (HQA) over semi-structured knowledge bases (SKBs) is vital for high-stakes domains like clinical precision medicine. However, HQA over SKBs is challenging because evidence is split across documents and relations, while answers must satisfy both semantic and structural constraints. Standard RAG pipelines retrieve evidence from textual or relational sources separately, failing to jointly capture the semantic conditions and relational constraints required by HQA over SKBs. Meanwhile, advanced agentic RAG systems that rely on computationally intensive models incur prohibitive inference latency and cost. To address these limitations, we introduce RSA-RAG, a Role-Specific Agentic RAG method with an asymmetric hybrid-path reasoning framework. We propose Role-Specific Asymmetric Collaboration for Hybrid-Path Reasoning, which assigns high-level roadmap planning and final answer generation to the large model, while lightweight role-specific agents handle frequent retrieval-control operations. Additionally, we introduce Strategy-Guided Candidate Navigation for Semi-Structured Retrieval, where compact agents perform strategy routing, constraint extraction, and hybrid navigation over an initial candidate set to retrieve textual and relational evidence around candidate targets. Furthermore, we implement Candidate-Support Evidence Structuring for Grounded Generation, which uses evidence verification to organize supported textual-relational evidence into structured grounding contexts for final generation. Extensive experiments show that RSA-RAG achieves strong retrieval performance on STARK, with an average MRR of 54.6% and competitive top-ranked performance across the three subsets. It further improves CRAG accuracy by 11.1% relative to Agentic RAG and reduces token consumption by up to 56.9% compared with implemented LLM-intensive baselines.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.