PolyUQuest Enhances Web RAG with Verifiable Structure-Aware Graph Retrieval

Current Retrieval-Augmented Generation (RAG) systems often treat web pages as flat text, overlooking the rich structural and semantic information embedded within HTML. To overcome this limitation, PolyUQuest has been introduced as a novel, verifiable, and structure-aware web RAG framework. This system is built upon a heterogeneous graph that seamlessly integrates hyperlink topology between pages, the Document Object Model (DOM) hierarchy within pages, and entity-relation knowledge across different web resources. PolyUQuest features a two-tier router that intelligently dispatches each user query to one of three specialized retrieval modes, depending on the query's structural requirements. These modes include direct block retrieval, cross-page graph traversal, and multi-hop entity reasoning. A key advantage of PolyUQuest is its full verifiability: every answer provided includes citations to its source page, heading path, and relevant entity links, allowing users to trace claims back to their original structural evidence. Evaluated on the Hong Kong Polytechnic University's official website, comprising thousands of pages and intricate data, PolyUQuest demonstrated superior performance compared to existing RAG systems. It achieved higher scores in answer correctness, coverage, and faithfulness, all while significantly reducing the number of LLM tokens consumed per query. An interactive demonstration further showcases its capabilities for inspecting answers and exploring evidence paths.