PaperlayerView exact revision on arXiv

Source-linked AI summary

PAGE-RAG: Provenance-Aware Graph Evidence Promotion for Fixed-Budget Multi-hop Retrieval-Augmented Generation

Haokun Deng, Xunkai Li, Hongchao Qin, Rong-Hua Li

arXiv:2608.29753v1cs.AI

TL;DR

Multi-hop RAG must recover complementary facts without filling a bounded reader context with related distractors, because connectivity alone does not ensure answer support. PAGE-RAG builds a query-local provenance-aware graph, scores support hypotheses, and minimally selects candidates; across benchmarks it improves support and answer F1 under the same final budget and also improves reported RAG backends.

  • Problem

    Multi-hop RAG faces a candidate-pool quality gap: expanded retrieval may contain needed facts, yet a bounded final context can remain dominated by related but non-supporting candidates.

  • Method

    PAGE-RAG builds a query-local provenance-aware graph over expanded candidates, scores paths by support signals, and selects a compact context for the reader.

  • Results

    PAGE-RAG improves support F1 and answer F1 by 10.4 and 3.3 points on a weighted average over a strong retriever and further improves all reported RAG backends.

  • Takeaways & Limitations

    PAGE-RAG provides a complete retrieval-to-reading pipeline and a backend-general promotion stage that improves context selection without replacing upstream retrieval logic.

  • Takeaways & Limitations

    PAGE-RAG cannot recover required support facts that never enter the upstream expanded candidate pool.

Abstract

from arXiv · show

Multi-hop question answering in retrieval-augmented gener?ation (RAG) often benefits from retrieving beyond the few candidates that will finally be read: narrow retrieval can miss an indispensable hop, while expanded retrieval introduces topical distractors. This challenge is not tied to a particu?lar knowledge-base format. Candidate pools may come from standalone retrievers, standard RAG backends, or graph-based retrieval pipelines. What is needed is a query-aware selection layer that can use relational structure to filter candidates be?fore generation. PAGE-RAG addresses this setting by using a graph as a temporary selection structure, rather than assum?ing a graph-structured knowledge base. It builds a query-local graph over retrieved candidates, records why candidates are connected, and treats each connection as a support hypothe?sis rather than support itself. We identify the resulting failure mode as a connectivity-support gap: connected candidates do not necessarily support the answer. We propose PAGE-RAG, a Provenance-Aware Graph Evidence promotion method that scores candidate paths with relevance, source-tracing meta?data, specificity, hubness, noise, and coherence signals, and applies minimal sufficient selection to promote supporting facts into a compact reader context. PAGE-RAG can serve as a complete retrieval-to-reading pipeline, and the same promo?tion stage can be inserted after existing retrieval or RAG sys?tems without replacing their upstream retrieval logic. Across three multi-hop QA benchmarks under the same final bud?get, PAGE-RAG improves support F1 and answer F1 by 10.4 and 3.3 points on a weighted average over a strong retriever. As a plug-in, PAGE-RAG further improves all reported RAG backends, including reasoning-oriented, compression-based, graph-based, and document/chunk-level systems.

Introduction

Multi-hop RAG must balance recovering missing hops against admitting distractors: connectivity among retrieved candidates does not guarantee answer support. PAGE-RAG addresses this gap with provenance-aware promotion that selects compact supporting context without replacing upstream retrieval systems.

  • Multi-hop answers require complementary facts to be retrieved together and placed in the right relation.
  • Expanded retrieval can recover indispensable hops but also introduces topical or graph-connected distractors that may degrade answers.
  • PAGE-RAG treats graph connections as support hypotheses and scores paths using query alignment, provenance, specificity, hubness, noise, and coherence signals.
  • PAGE-RAG promotes support-bearing candidates into a compact reader context while preserving the reachability gained from expanded retrieval.
  • PAGE-RAG can operate as a complete retrieval-to-reading pipeline or attach after existing retrievers and RAG systems without replacing upstream logic.

Preliminaries

The problem setting assumes an upstream knowledge-access module that returns an expanded candidate pool, while the reader can consume only a bounded subset. Existing retrieval and graph-based approaches expose useful relations but may still leave complementary facts mixed with distractors.

  • The upstream module may be a standalone retriever, standard RAG backend, or graph-based system, and the underlying knowledge base need not be a graph.
  • The reader receives at most k context units, so selection must find a compact subset that jointly supports the answer.
  • Ordinary top-k retrieval emphasizes local relevance, whereas graph retrieval may treat structural connectivity as useful even when candidates do not support the answer.
  • Standalone retrievers can expose semantically related candidates while missing complementary facts whose relevance emerges only after another hop.
  • Reasoning-oriented, selective, and graph-based RAG systems use reasoning, reflection, or relational structure, but graph links can still introduce noisy or non-supporting bridges.

Method

PAGE-RAG expands retrieval into a query-local provenance-aware graph, evaluates connected paths as support hypotheses, and selects a minimal context under the fixed reader budget. The design is temporary, backend-general, and focused on support rather than raw connectivity.

  • Expanded Candidate Retrieval: PAGE-RAG retrieves a broad candidate pool so missing multi-hop facts become reachable without enlarging the reader input.
  • Provenance-Aware Evidence Graph: The expanded pool is converted into a per-question graph that exposes relational structure while remaining a temporary selection workspace.
  • Provenance-Aware Evidence Graph: Each graph edge records source-tracing metadata explaining why candidates are connected, including identifiers, triggers, confidence, diversity, specificity, and hubness.
  • Support-Aware Path Scoring: Candidate paths are scored from sentence support, edge support, coherence, length cost, reliability, relevance, and noise rather than connectivity alone.
  • Ranking and Minimal Selection: Minimal selection chooses complementary paths that add support while accounting for redundancy and reader-budget cost.
  • Backend Generalization: The promotion stage can follow existing retrievers or RAG systems and preserves their reader interface while producing a compact context.

Experiments

Experiments evaluate PAGE-RAG as both a complete retrieval-to-reading pipeline and a plug-in under fixed reader budgets, showing stronger answer and support quality across retrieval and RAG settings. Ablations and ranking analyses attribute these gains to support-aware promotion and minimal selection, while results also identify dependence on upstream candidate pools.

  • Evaluation setup: Experiments test PAGE-RAG both as a complete retrieval-to-reading workflow and as a plug-in between candidate generation and reading.The evaluations assess final-context quality without increasing the final reader input.
  • Evaluation setup: The sentence-level protocol fixes the reader input to five sentences, while the mixed-granularity protocol compares document, passage, or chunk systems through paired baseline/plugin comparisons.The evaluation includes standalone retrievers, graph-based retrieval, and reasoning or compression-oriented RAG systems.
  • Complete pipeline results: 10.45 Ans F1 and 10.45 Sup F1 are the gains from NV-Embed-v2 to PAGE-RAG, rising from 58.00 to 65.00 Ans F1 and from 44.86 to 55.31 Sup F1.PAGE-RAG also outperforms IRCoT@5 and RECOMP@5 in both answer and support F1, achieving the best answer F1 in the sentence-level table.
  • Plug-in results: 14.42 Ans F1 and 18.33 Sup F1 are PAGE-RAG's improvements over IRCoT@5 on MuSiQue, while all six reported document/chunk-level backends also improve.The plug-in gains span multi-step retrieval, compression, self-reflective, efficient, and graph-based RAG systems.
  • Scope: PAGE-RAG depends on the expanded upstream candidate pool and cannot recover support facts that never enter that pool.Its gains are generally smaller on graph-based backends because those systems already expose relational structure before promotion.
  • Ranking analysis: PAGE-RAG reduces Out support from 31.5% to 27.1% versus NV-Embed-v2 and shifts more gold supporting facts into visible top-5 positions.In plug-in settings, Out support falls from 33.4% to 28.0% for IRCoT and from 15.9% to 4.7% for SelfRAG.
  • Ablation: Removing support-aware scoring lowers average Ans F1 to 50.64, while removing minimal selection lowers it to 54.44, demonstrating that both components are necessary.The ablation tests answer F1 across all three datasets.

Conclusion

The conclusion frames PAGE-RAG as evidence promotion that separates graph reachability from answer support. Across evaluation settings, it improves retrieval-to-reading and plug-in systems by ranking support-bearing facts into compact contexts, while requiring both support-aware scoring and minimal selection.

  • Conclusion: PAGE-RAG separates candidate reachability from final-context usefulness by scoring paths for support and selecting a compact context before generation.The framework uses a query-local support graph and provenance-aware evidence promotion.
  • Conclusion: Experiments show improvements in complete retrieval-to-reading pipelines and plug-in settings over existing RAG and GraphRAG backends.Ranking analysis shows more gold supporting facts enter the final visible context.
  • Conclusion: Ablation results confirm that support-aware scoring and minimal selection are both necessary for the observed improvement.
  • Conclusion: The method cannot recover facts that never enter the upstream candidate pool.

Code Availability

The original PAGE-RAG implementation is publicly available under the Apache License 2.0, while third-party baseline components retain their original licenses.

  • Code Availability: The original PAGE-RAG implementation is available on GitHub under the Apache License 2.0.Third-party baseline components retain their original licenses.

Appendix A: Method Details

The appendix outlines the PAGE-RAG workflow of expanding retrieval, constructing a query-local graph, scoring paths, and selecting a fixed-budget reader context.

  • PAGE-RAG expands the candidate pool before building a query-local graph for selection.The graph construction and support-path scoring are part of the high-level workflow.
  • Table 4 summarizes the graph edge types used in the query-local PAGE-RAG graph.

A.1 Query-Local Graph Construction

PAGE-RAG constructs a query-local graph over retrieved candidates and treats graph connections as support hypotheses rather than evidence by default. It scores provenance-aware paths and selects a minimal sufficient context under a fixed budget.

  • Graph construction: PAGE-RAG builds a temporary graph only over candidates retrieved for the current question, without requiring a corpus-wide graph index.The graph includes document or passage, sentence, and entity nodes.
  • Graph construction: Document–sentence and sentence–entity edges preserve source traceability, while sentence–sentence edges provide the main support-promotion paths.The graph contains five edge types summarized in the paper’s Table 4.
  • Support hypotheses: Graph edges are proposals for support, not support by default, operationalizing PAGE-RAG’s connectivity-support gap.This distinguishes PAGE-RAG’s graph use from graph-based RAG systems that organize text through relational links.
  • Feature scoring: PAGE-RAG records edge provenance, triggers, confidence, source diversity, specificity, hubness, noise, and question-conditioned relevance for path scoring.Shared-entity edges derive triggers from intersecting sentence entities, while edge scores combine relevance, reliability, specificity, hubness, and noise.
  • Path scoring: Path scores combine sentence relevance, edge scores, path length, multi-hop structure, and coherence, with a bounded bridge bonus for plausible question-answer transitions.The bridge bonus applies when a path contains at least two sentences, including a question entity and a relation signal.
  • Context selection: Minimal sufficient selection adds a path only when it improves support utility or state, preserves the budget, and may retain one plausible intermediate bridge before rejection.Afterward, unused slots are filled with the highest-scoring remaining candidates to keep the final reader budget fixed in controlled experiments.

A.6 Document and Chunk-Level Adaptation

PAGE-RAG adapts its sentence-level promotion principle to document-, passage-, and chunk-level backends. It preserves each backend’s native output format while using finer internal structure when available.

  • Document and chunk-level adaptation: PAGE-RAG can apply its promotion principle to document-, passage-, or chunk-level retrieval backends.When text structure is available, it still builds sentence-level or chunk-internal paths for more precise support scoring.
  • Document and chunk-level adaptation: The final promoted context is mapped back to the backend’s native unit, such as a document or chunk identifier, while provenance features are interpreted at that granularity.Source diversity becomes diversity across documents or chunks, and selected units must remain within the final top-k budget.
  • Plug-in integration: As a plug-in layer, PAGE-RAG accepts candidate pools from upstream backends, recalibrates them with a temporary graph, and returns the context-unit type expected by the reader.This insertion does not require replacing the upstream retrieval logic.

Appendix B: Feature-Group Ablation

The ablation evaluates four interpretable support-scoring families and finds that their contributions vary by dataset, while the full scorer outperforms feature-group removals.

  • Feature groups: PAGE-RAG groups support-scoring signals into query alignment, provenance reliability, bridge specificity, and path quality.These families address question alignment, source traces, distractor-prone bridges, and support-path plausibility.
  • Dataset differences: Different datasets stress different scorer components because missing support appears as explicit bridges, paragraph-level associations, or named-entity relation chains.HotpotQA emphasizes bridge entities, MuSiQue paragraph-level associations, and 2Wiki relation chains.
  • Query alignment: 14.84 Ans F1 points: removing query alignment drops HotpotQA from 70.94 to 56.10 Ans F1.The passage also reports a drop on 2Wiki from 63.23 to 44.78 Ans F1.
  • Query alignment: 18.45 Ans F1 points: removing query alignment drops 2Wiki from 63.23 to 44.78 Ans F1.Query alignment is especially important where questions require named-entity relation chains or entity-bearing bridge sentences.
  • Bridge specificity and path quality: Removing bridge specificity causes the largest answer drop on MuSiQue, while path quality also helps separate compact support chains from distracting associations.This reflects MuSiQue’s paragraph-style questions and generic-entity or long-association distractors.
  • Provenance reliability: Provenance reliability contributes steadily across datasets by separating clearer source traces from weaker or less repeated links, especially when relevance and graph structure are ambiguous.The ablation does not claim provenance alone solves support promotion; it stabilizes the scorer under ambiguity.
  • Combined scorer: The four signal families are complementary: removing any one leaves a partial view of support, and the full scorer is substantially stronger than group removals.The result supports calibrating connectivity with multiple signal families rather than treating it as answer support by default.

Appendix C: Noise Robustness

The noise stress test appends distractor documents to the expanded pool while fixing the reader budget. PAGE-RAG degrades gradually rather than collapsing, but performance declines monotonically as noise increases.

  • Setup: The stress test appends +5, +10, or +20 distractor documents before graph construction while keeping the final reader context fixed at five units.All runs use the same DeepSeek-V4-Pro reader as the main table.
  • Results: HotpotQA answer F1 drops 2.68, 4.07, and 5.26 points with +5, +10, and +20 distractors, respectively.The reported metric is answer F1 under the fixed final reader budget.
  • Results: MuSiQue answer F1 drops 4.93, 5.43, and 6.62 points with +5, +10, and +20 distractors, respectively.The decline is measured as additional distractors are appended to the expanded candidate pool.
  • Results: 2Wiki answer F1 drops 5.65, 6.97, and 7.86 points with +5, +10, and +20 distractors, respectively.The 2Wiki reductions are the largest among the three reported benchmarks at every tested noise level.
  • Interpretation: The noisy PAGE-RAG variants remain stronger than, or roughly comparable to, common strong baselines under the same final budget.The authors interpret the experiment as sensitivity analysis rather than noise-invariant robustness.

Appendix D: Case Study

The case studies show PAGE-RAG recovering missing supporting facts already present in an expanded top-20 pool and promoting them into the final top-5 context. Across three examples, this changes initially incorrect answers without increasing the reader budget.

  • Case-study setup: The case studies examine three questions where initial top-5 retrieval misses required support, but expanded top-20 retrieval contains the missing facts.Tables 8 and 9 summarize answer changes and the movement of facts into final context.
  • HotpotQA: HotpotQA changes from “Slovenia” to “Phoenix Suns” after PAGE-RAG promotes nationality and season facts from ranks 6, 8, and 12.The initial retrieval is dominated by basketball-position sentences and lacks the facts connecting Goran Dragic to the 2013-2014 Phoenix Suns season.
  • MuSiQue: MuSiQue changes from “1970” to the correct date after PAGE-RAG promotes the Majuro location chain from ranks 13 and 15.The promoted facts state that Tobolar Copra is in Majuro and that Majuro’s first mosque opened in September 2012.
  • 2Wiki: 2Wiki changes from Ibrahim’s father to the correct grandfather after PAGE-RAG promotes both missing paternal-chain links from ranks 6 and 19.The two promoted facts connect Ibrahim Shah to Raja Ibrahim bin Raja Lumu and Salehuddin Shah to Raja Lumu bin Daeng Chelak.
  • Cross-case pattern: Across the examples, PAGE-RAG selects complementary connected paths so buried missing hops enter the final reader context without increasing the reader budget.The expanded pool already contains the needed facts, while topical distractors obscure them from the initial retrieval.
Loading 2608.29753v1…