The DOM Block Becomes the Citation Receipt
A July 2026 arXiv paper turns an institutional website into a graph of pages, DOM blocks, entities, and topics. The governance issue is not whether retrieval gets smarter. It is whether a generated answer can show the exact page structure that made each claim possible.
The Paper
The paper is Ying Liu, Yi Ye, Quanyu Feng, Mingxi Ye, Mingtao Zhang, Haoyang Li, Chen Jason Zhang, and Qing Li's PolyUQuest: Verifiable Structure-Aware Web RAG over Heterogeneous Graphs, arXiv:2607.08269 [cs.AI]. The arXiv API lists version 1 as submitted on July 9, 2026. The PDF metadata reports 5 pages, and the title page lists The Hong Kong Polytechnic University in Hong Kong SAR, China.
This belongs beside the site's work on retrieval-augmented generation, AI search and answer engines, curated public-answer corpora, citation verification, and web-agent evidence trails. Its fresh angle is structural provenance: which page, heading, DOM block, and graph path carried the claim.
What It Builds
PolyUQuest models a website as a heterogeneous graph with four node types: webpages, evidence blocks, entities, and topics. Its edge set records page-to-page hyperlinks, page-to-block containment, block-to-entity mentions, and semantic associations among entities and topics. A block carries its source page, heading path, embedding, length, and mentioned entities, so retrieval output is tied to an inspectable part of the site rather than an anonymous text chunk.
The offline pipeline cleans HTML, removes scripts, styles, frames, comments, repeated navigation, footers, and cookie notices, then builds heading-aware evidence blocks. Tables, lists, and preformatted code are treated as atomic units. Online, a two-tier router sends a query to one of three retrieval modes: Mode A for direct factual lookup, Mode B for navigation across linked pages, and Mode C for entity-based multi-hop reasoning. When heuristics are not enough, an LLM classifier assigns the mode with a confidence score.
The Benchmark Signal
The demo evaluation uses 4,240 official English PolyU pages covering programme, staff, department, and research sections. The built graph contains 31,086 DOM blocks, 29,119 resolved entities, and 37,680 relations. The paper evaluates 300 PolyU-Web questions covering single-page factual lookup, cross-page navigation, and entity-centric reasoning. For each question, the authors manually annotate the relevant official PolyU pages and reference evidence blocks.
Against ChunkRAG, HtmlRAG, FastGraphRAG, and LightRAG, with the same generator LLM, prompt, and embedding model, PolyUQuest reports the strongest table results: answer correctness 0.644, coverage 0.649, and faithfulness 0.921. It reports 2,968 average LLM tokens per query, close to ChunkRAG's 2,947 and below HtmlRAG's 4,009, FastGraphRAG's 4,484, and LightRAG's 29,825. The ablation table says replacing DOM blocks with fixed-size chunks lowers correctness to 0.563 and coverage to 0.510; removing cross-page mode lowers correctness to 0.624 and coverage to 0.633.
Why It Matters
Most institutional websites are not one document. They are nested pages, menus, course lists, staff biographies, policy pages, forms, calendars, and repeated names. A flat RAG system can retrieve plausible chunks while losing why one page should be read as a subpage, prerequisite, office contact, programme requirement, or obsolete neighbor. PolyUQuest's useful move is to treat site structure as part of the evidence.
That matters for public service and campus answer engines. A student asking about requirements, a patient asking about a clinic pathway, or a citizen asking about an agency service needs more than a fluent paragraph and a URL. The answer should preserve whether a claim came from a current programme page, a heading under admission rules, a staff profile, or an entity link across several official pages. Structure is not decoration; it is part of institutional meaning.
What It Does Not Prove
The result is still a demo paper, not a universal answer-engine certificate. The evaluation is on one university website, one 300-question benchmark, and one institutional domain. The authors say porting to another domain requires a fresh crawl and a domain-specific entity schema. That is not a small detail for hospitals, governments, or firms whose entity types, legal authority, retention rules, and page freshness signals differ sharply.
The official GitHub repository also narrows the reproducibility claim. Its README says the repository contains the online serving stack: a FastAPI backend and Next.js frontend using Neo4j and Qdrant. The stores start empty, the offline indexing pipeline is not included, code is Apache-2.0 licensed, and the dataset is for research use only. The paper is strongest as evidence for a provenance architecture and reported benchmark, not as proof that any deployed site can reproduce the full pipeline unchanged.
Governance Reading
The Spiralist concern is citation without location. A link to a page is often too coarse. A page can contain current policy, outdated notes, a contact box, a footnote, a table, a heading that scopes the claim, and a navigation shell that should not be read as content. When a model cites the whole page, the user is asked to trust that the right part of the page was used.
PolyUQuest points toward a stricter answer contract. A citation should be a receipt with source URL, page title, heading path, evidence block, entity links, retrieval mode, and graph path. That does not make the generated answer true by itself. It makes the support inspectable. It lets a reviewer ask whether the retrieved block actually covers the claim, whether the heading changes its scope, whether a cross-page traversal skipped a more authoritative page, and whether an entity merge created a false identity.
The Receipt
A structure-aware RAG receipt should include the crawl date, crawl boundary, excluded page patterns, graph schema, entity schema, entity-resolution rule, block-construction rule, boilerplate filters, source URL, page title, heading path, DOM block text or hash, cited entities, retrieval mode, router confidence, graph traversal path, prompt, model versions, embedding model, generator model, query tokens, build tokens, answer text, unsupported-claim flags, and reviewer decision.
The practical rule: once a website answer comes from structured retrieval, structure becomes part of accountability. A citation without its block and path is not yet a receipt. It is only a doorway.
Sources
- Ying Liu, Yi Ye, Quanyu Feng, Mingxi Ye, Mingtao Zhang, Haoyang Li, Chen Jason Zhang, and Qing Li, PolyUQuest: Verifiable Structure-Aware Web RAG over Heterogeneous Graphs, arXiv:2607.08269 [cs.AI], submitted July 9, 2026.
- arXiv API record for arXiv:2607.08269, checked for title, authors, subject class, submission date, update date, and abstract.
- arXiv PDF for arXiv:2607.08269, checked for page count, affiliations, graph design, retrieval modes, evaluation setup, reported results, ablations, generalizability discussion, and GenAI disclosure.
- arXiv experimental HTML for PolyUQuest, checked for system details, source-block provenance fields, demo interface, tables, and conclusion.
- Official GitHub repository 13-pieces-teen/PolyUQuest, checked for serving-stack release, README limitations, dependency stack, license note, and dataset note.