PaperlayerView exact revision on arXiv

Source-linked AI summary

MOSAIC: Query-Aware Exploration Policy Adaptation for GraphRAG

EunKyeong Lee, Kyeong-Jin Oh, Jinwon Kim, Hye Woo Lee, Minsang Song, Hyeongjun Jang, Junyoung Youn

arXiv:2609.11065v1cs.AI

TL;DR

GraphRAG systems often apply shared exploration settings despite queries requiring different evidence structures. Mosaic uses an LLM analyzer to construct bounded per-query retrieval policies, improving answer correctness and evidence recall while reducing graph search relative to uniformly wide exploration.

  • Problem

    GraphRAG retrieval must decide where to start, how to traverse, when to stop, and which paths to preserve, but globally shared policies fit heterogeneous questions poorly.

  • Method

    Mosaic uses a training-free analyzer to convert query-specific evidence requirements into bounded controls for seed selection, traversal, stopping, and evidence retention while keeping the graph and answer generator shared.

  • Results

    76.97% overall ACC on Medical and 64.33% on Novel, with Evidence Recall of 95.1% and 90.2%; Mosaic also outperforms fixed exploration scopes on Medical.

  • Takeaways & Limitations

    Effective GraphRAG should adapt the operating policy of graph exploration itself, not merely relevance scores, while using substantially less graph search than uniformly wide exploration.

  • Takeaways & Limitations

    The analyzer can misclassify evidence structure, and transfer to different graph schemas may require additional calibration.

Abstract

from arXiv · show

Graph Retrieval-Augmented Generation (GraphRAG) can connect evidence distributed across a corpus graph, but most systems use largely shared exploration procedures across queries. This creates a structural mismatch: direct facts may need compact local neighborhoods, comparisons need balanced coverage of multiple targets, and mediated questions may require deeper paths through weakly related connectors. We present Mosaic, a training-free framework that formulates GraphRAG retrieval as a per-query control problem. An LLM analyzer converts query-specific evidence requirements into a bounded policy over seed selection, graph traversal, stopping, and evidence selection, while the corpus graph, indexes, scoring functions, grounding procedure, and answer generator remain shared. On GraphRAG-Bench, Mosaic achieves query-weighted Answer Correctness of 76.97 on Medical and 64.33 on Novel, improving over the strongest previously reported overall results by 5.13 and 4.43 points. On Medical, it reaches 95.1 Evidence Recall and 86.1 Context Relevancy. Controlled comparisons on an identical graph and generator show that no fixed narrow, medium, or wide policy is consistently optimal; Mosaic improves by 9.96 points over the strongest canonical fixed policy. Relative to Fixed Wide, it evaluates 81.9% fewer paths and retains 47.2% fewer evidence items. Transfer experiments on HotpotQA, MuSiQue, and 2WikiMultiHopQA further show that the policy interface can be applied without benchmark-specific retriever training.

1 Introduction

GraphRAG retrieval often applies shared exploration limits despite questions requiring different evidence structures. Mosaic constructs bounded per-query policies while keeping the graph and answer generation shared, improving accuracy and reducing search relative to fixed policies.

  • 1 Introduction: Mosaic translates query-specific evidence requirements into bounded controls for seeds, traversal, stopping, and evidence retention while sharing the corpus graph and answer generator.The policy interface is training-free and extensible to new evidence requirements.
  • 1 Introduction: No fixed exploration scope is uniformly effective: Fixed Medium reaches 67.01% ACC on Medical, versus 65.45% for narrower and 66.35% for wider exploration.
  • 1 Introduction: 81.9% fewer paths are evaluated and 47.2% fewer evidence items retained relative to Fixed Wide.The reduction accompanies the reported gains rather than exhaustive search.
  • 1 Introduction: Connector traversal is activated for 65.6% of multi-hop queries but 21.6% of other queries, indicating requirement-aligned policy variation.
  • 1 Introduction: The framework transfers to three standard multi-hop QA benchmarks without benchmark-specific retriever training.

2 Related Work

Prior GraphRAG systems adapt retrieval in several ways, but Mosaic distinguishes itself by constructing an explicit, bounded operating policy across retrieval stages within one shared pipeline.

  • 2 Related Work: GraphRAG research includes community-oriented search, entity- and relation-level retrieval, personalized PageRank propagation, flow-based path pruning, learned graph retrievers, and adaptive evidence operators.
  • 2 Related Work: The adaptation interface is positioned against systems whose operating choices remain shared at inference time.Table 1 is described as comparing where query adaptation occurs across representative systems.
  • 2 Related Work: Mosaic jointly controls seeds, traversal, stopping, and retained evidence without training an additional router or graph retriever.Its objective is query-specific policy construction rather than merely query-conditioned scoring or predefined query classification.
  • 2 Related Work: The formulation replaces a globally selected policy with a bounded query-specific policy while keeping the graph, retrieval implementation, and answer model shared.The analyzer generates the policy only from the question, with validation, clamping, and conservative defaults.
  • 2 Related Work: Query-specific policy construction translates a query into composable stage-level controls executed by a common pipeline.Mosaic implements this third form of adaptation; diagnostic labels do not select separate retrievers.

4 From Retrieval Failures to Policy Signals

Mosaic derives policy signals from recurring retrieval failures rather than imposing mutually exclusive question types. Its signals are extensible, and controls may interact within a single query.

  • 4 From Retrieval Failures to Policy Signals: Recurring failures expose six questions for configuring retrieval, rather than six mutually exclusive question types.The current signals summarize observed failures and guide the analyzer’s policy construction.
  • 4 From Retrieval Failures to Policy Signals: The six signals are not a closed taxonomy; new failure patterns can add reasoning cues mapped to existing or newly introduced bounded controls.The analyzer-to-policy interface, not the number or names of current cues, is the core contribution.
  • 4 From Retrieval Failures to Policy Signals: Controls are coupled rather than independent treatments: a comparison may require per-target seeds, connector traversal, and a larger evidence budget together.Control-specific counterfactuals are therefore diagnostic rather than an additive decomposition of total performance.

5 MOSAIC

Mosaic formulates GraphRAG retrieval as a validated, query-specific policy over shared seed, traversal, stopping, and evidence-selection stages. Its bounded controls combine structured path scoring, connector exploration, cumulative-gain stopping, role-aware evidence selection, and source-window reranking before fixed answer generation.

  • End-to-end architecture: Mosaic maps each query to a validated policy configuring seed selection, traversal, stopping, and evidence controls while keeping answer generation fixed.The pipeline uses a shared graph, retrieval process, and answer model; the analyzer changes retrieved evidence rather than free-form answer generation.
  • Policy space and validation: The analyzer starts from conservative defaults and emits bounded controls whose unsupported values are rejected, numeric fields clamped, and missing values reverted.This finite interface prevents direct arbitrary-code execution or invention of new retrieval operators.
  • Seed selection: Seed selection combines entity- and relation-index scores, with per-target mode preserving candidates for each target before deduplication and truncation.Per-target preservation prevents comparisons from allocating all seeds to their most salient entity.
  • Policy-conditioned traversal: Structured path scoring combines relevance, target, specificity, anchor, transition, coherence, hub, repetition, and question-like-label signals, with adaptive weights for anchored queries.At each depth, retained paths are expanded, connector candidates may be added, and the best Wq paths are kept under hard cap Dq.
  • Policy-conditioned traversal: Connector exploration searches paths joining regions initialized by distinct targets, while standard and connector views share the same scoring and pruning pool.This adds a second exploration view without creating a separate retrieval pipeline.
  • Cumulative-gain stopping: Mosaic stops after the minimum depth when novelty gain is low and evidence completeness reaches 0.35, or when the query-specific depth cap is reached.The implementation uses a shallow, medium, or deep threshold and patience-one stopping; selection then preserves core and complementary roles before score-based filling.
  • Evidence selection: After path selection, deduplicated evidence is capped and restored to a preservation floor, while local source-window reranking supplies detailed textual context for answer generation.The evaluated configuration does not add a separate post-traversal LLM noise-filtering call.

6 Experimental Setup

The evaluation uses all questions from Medical and Novel GraphRAG-Bench across four task categories, with query-count-weighted Answer Correctness as the primary end-to-end metric. Mosaic is compared both with externally reported systems and with fixed policies on an identical graph and generator.

  • Primary benchmarks: GraphRAG-Bench Medical contains 2,062 questions over 2,406 medical-guideline documents, while Novel contains 2,010 questions over 461 literary documents.Both benchmarks include Fact Retrieval, Complex Reasoning, Contextual Summarization, and Creative Generation, using every question and the original corpora.
  • Metrics: Answer Correctness is the primary end-to-end metric, computed as query-count-weighted overall ACC; Evidence Recall and Context Relevancy evaluate retrieval.Task-specific ROUGE-L, coverage, and faithfulness are secondary metrics.
  • Implementation and controls: Mosaic and fixed-policy controls share each corpus graph, indexes, retrieval code, and answer generator, and Mosaic requires no additional training or fine-tuning.Graphs are built once with a LightRAG-style extraction pipeline, while OpenAI text-embedding-3-large and GPT-4o-mini support representations, analysis, and generation.
  • Baselines: Published-system comparisons use externally reported values, while Fixed Narrow, Medium, Wide, and Budget-Matched Fixed provide controlled comparisons on the identical graph and generator.This separates broad benchmark comparison from the within-implementation test of policy adaptation.

7 Results

Mosaic improves accuracy and retrieval quality over fixed policies while using less graph-search effort, though its analyzer call increases end-to-end latency.

  • 7.1 Answer correctness: 76.97 on Medical and 64.33 on Novel improve over the strongest previously reported overall results by 5.13 and 4.43 points.
  • 7.2 Retrieval quality: Mosaic achieves the highest Evidence Recall on both domains and the highest reported Medical Context Relevancy.It exceeds G-Reasoner by 1.3 points on Medical and 2.5 points on Novel for Evidence Recall.
  • 7.3 Why one fixed policy is insufficient: Mosaic exceeds the strongest canonical fixed policy by 9.96 points, showing that query-level allocation outperforms choosing one fixed operating point.Medium improves on Narrow, while Wide declines despite broader search and more evidence.
  • 7.4 Analyzer behavior: Across 2,062 Medical queries, the analyzer produces 19 policy combinations rather than collapsing to one configuration.Assigned depth spans 2, 3, and 4 hops, while seed counts range from 5 to 12 and evidence budgets from 8 to 16.
  • 7.5 Diagnostic counterfactuals: Diagnostic counterfactuals are non-additive because the same query can improve through several interacting controls.The analysis identifies output cardinality as the strongest single count, consistent with high-scoring paths crowding out sibling evidence for list-like answers.
  • 7.6 Latency and graph-search effort: Mosaic is not faster end-to-end: its policy-construction call adds 2.82 seconds, yielding 9.68 seconds versus 5.29 for Fixed Medium.The efficiency benefit is reduced graph search rather than lower total latency.

8 Transfer to Multi-Hop QA

Mosaic transfers its policy interface to three multi-hop QA benchmarks without benchmark-specific retriever training, with mixed results across datasets.

  • 8 Transfer to Multi-Hop QA: The transfer experiments use 1,000-question subsets without benchmark training splits or task-specific retriever fine-tuning.Only one or two short answer-format instructions are added per benchmark.
  • 8 Transfer to Multi-Hop QA: Mosaic is close to the trained reference on HotpotQA and exceeds it by 1.8 EM on MuSiQue, while a substantial gap remains on 2WikiMultiHopQA.
  • 8 Transfer to Multi-Hop QA: Semantic correctness scores of 77.9, 51.3, and 75.1 suggest that lexical EM/F1 penalize some semantically correct answers with different formatting.The comparisons do not establish state-of-the-art transfer because systems were not rerun under a common pipeline and G-Reasoner uses benchmark-specific supervision.

9 Qualitative Analysis

Qualitative cases show that combined controls can recover difficult evidence paths, but expressive query-specific control can still overreact and misinfer evidence requirements.

  • 9 Qualitative Analysis: More depth alone does not recover the kidney-tumor answer when nearby but incorrect staging evidence dominates the larger context.
  • 9 Qualitative Analysis: Guardrails constrain policy outputs but cannot guarantee that the analyzer infers the evidence requirement correctly.
  • 9 Qualitative Analysis: Transfer results are summarized in Table 10, where G-Reasoner values are externally reported and use target-benchmark training.

10 Discussion

The results support query-specific policy adaptation as a practical shared-pipeline design, while leaving individual prompt effects and component-matched comparisons unresolved.

  • 10.1 What the results establish: Controlled experiments show that query-specific policy adaptation improves over several globally fixed operating points on a shared graph and generator.
  • 10.1 What the results establish: Policy logs indicate that the controller neither collapses to one configuration nor varies arbitrarily, while graph-search measurements show gains do not come from uniformly retrieving more evidence.
  • 10.2 What the results do not establish: The experiments do not isolate causal contributions for every prompt phrase or low-level scoring term, and control-specific counterfactuals are non-additive.
  • 10.2 What the results do not establish: Comparisons to published GraphRAG systems are not component-matched reproductions, and transfer comparisons use asymmetric training conditions.
  • 10.1 What the results establish: The bounded policy schema separates safe retrieval controls from analyzer decisions and monitoring, allowing new failure patterns to be addressed without a new end-to-end pipeline for each query type.

11 Limitations

Mosaic has important limitations involving analyzer reliability, transferability, graph quality, cost, and evaluation scope.

  • The LLM analyzer can misclassify evidence structure, as demonstrated by the SLNB case.
  • Current signals and valid ranges were developed for the evaluated setting, so different graph schemas may require additional calibration.
  • Mosaic’s current implementation has higher latency and monetary cost than fixed retrieval, while some baseline comparisons remain externally reported.

12 Conclusion

Mosaic frames GraphRAG retrieval as per-query control rather than globally fixed exploration. Its query-specific bounded policies improve answer correctness and evidence recall while using less graph search than uniformly wide retrieval.

  • Mosaic treats GraphRAG retrieval as a per-query control problem with bounded policies for seeds, traversal, stopping, and evidence retention.
  • The framework improves overall answer correctness and evidence recall on GraphRAG-Bench while using substantially less graph search than a uniformly wide policy.
  • The six current considerations are composable implementation signals rather than a fixed taxonomy, supporting analyzer-driven query-specific exploration as an extensible design principle.
Loading 2609.11065v1…