Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Navigational Approach for Comprehensive RAG via Traversal over Proposition Graphs

Maxime Delmas, Lei Xu, André Freitas

TL;DR

ToPG addresses the limitations of standard RAG by uniting proposition-level granularity with graph connectivity through a heterogeneous proposition graph. It introduces Suggestion-Selection cycles and three operational modes (Naive, Local, Global) to support factoid, multi-hop, and abstract QA. Across experiments on Simple, Complex, and Abstract QA tasks, ToPG achieves competitive or superior results compared to structure-augmented baselines, with notable gains in multi-hop and abstract reasoning, and provides insights into token-cost trade-offs. The approach demonstrates that query-aware graph traversal over dense factual units is a key driver for scalable and capable structured RAG systems.

Abstract

Standard RAG pipelines based on chunking excel at simple factual retrieval but fail on complex multi-hop queries due to a lack of structural connectivity. Conversely, initial strategies that interleave retrieval with reasoning often lack global corpus awareness, while Knowledge Graph (KG)-based RAG performs strongly on complex multi-hop tasks but suffers on fact-oriented single-hop queries. To bridge this gap, we propose a novel RAG framework: ToPG (Traversal over Proposition Graphs). ToPG models its knowledge base as a heterogeneous graph of propositions, entities, and passages, effectively combining the granular fact density of propositions with graph connectivity. We leverage this structure using iterative Suggestion-Selection cycles, where the Suggestion phase enables a query-aware traversal of the graph, and the Selection phase provides LLM feedback to prune irrelevant propositions and seed the next iteration. Evaluated on three distinct QA tasks (Simple, Complex, and Abstract QA), ToPG demonstrates strong performance across both accuracy- and quality-based metrics. Overall, ToPG shows that query-aware graph traversal combined with factual granularity is a critical component for efficient structured RAG systems. ToPG is available at https://github.com/idiap/ToPG.

A Navigational Approach for Comprehensive RAG via Traversal over Proposition Graphs

TL;DR

ToPG addresses the limitations of standard RAG by uniting proposition-level granularity with graph connectivity through a heterogeneous proposition graph. It introduces Suggestion-Selection cycles and three operational modes (Naive, Local, Global) to support factoid, multi-hop, and abstract QA. Across experiments on Simple, Complex, and Abstract QA tasks, ToPG achieves competitive or superior results compared to structure-augmented baselines, with notable gains in multi-hop and abstract reasoning, and provides insights into token-cost trade-offs. The approach demonstrates that query-aware graph traversal over dense factual units is a key driver for scalable and capable structured RAG systems.

Abstract

Standard RAG pipelines based on chunking excel at simple factual retrieval but fail on complex multi-hop queries due to a lack of structural connectivity. Conversely, initial strategies that interleave retrieval with reasoning often lack global corpus awareness, while Knowledge Graph (KG)-based RAG performs strongly on complex multi-hop tasks but suffers on fact-oriented single-hop queries. To bridge this gap, we propose a novel RAG framework: ToPG (Traversal over Proposition Graphs). ToPG models its knowledge base as a heterogeneous graph of propositions, entities, and passages, effectively combining the granular fact density of propositions with graph connectivity. We leverage this structure using iterative Suggestion-Selection cycles, where the Suggestion phase enables a query-aware traversal of the graph, and the Selection phase provides LLM feedback to prune irrelevant propositions and seed the next iteration. Evaluated on three distinct QA tasks (Simple, Complex, and Abstract QA), ToPG demonstrates strong performance across both accuracy- and quality-based metrics. Overall, ToPG shows that query-aware graph traversal combined with factual granularity is a critical component for efficient structured RAG systems. ToPG is available at https://github.com/idiap/ToPG.
Paper Structure (44 sections, 11 equations, 10 figures, 6 tables)

This paper contains 44 sections, 11 equations, 10 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Graph Construction
  4. Graph Navigation: Suggestion–Selection Cycles
  5. Suggestion-Selection Retrieval.
  6. Query and Graph Aware Suggestions
  7. Subgraph Extraction
  8. Naive, Local and Global modes
  9. Naive
  10. Local
  11. Global
  12. Steps 1-2: Seeding
  13. Step 3: Compute Queries
  14. Step 4: Iteratively explore and collect
  15. Step 5: Identifying communities
  16. ...and 29 more sections

Figures (10)

  • Figure 1: The ToPG framework. The system operates on a heterogeneous graph where propositions connect entities and passages. ToPG navigates this graph using iterative Suggestion-Selection cycles, allowing for three operational modes: Naive (factoid retrieval), Local (multi-hop inference), and Global (community-based search).
  • Figure 2: On the left, panel Local shows a step-by-step example for the Local mode. On the right panel, step-by-step description of the Global mode.
  • Figure 3: Win rates (%) of ToPG against GraphRAG and LightRAG across 3 corpora and 4 criteria, with increasing number of collected propositions ($200$-$1000$) and w/ or w/o Rocchio-style feedback.
  • Figure 4: $\text{Log}_{10}(\text{Token-counts})$ between baselines evaluated on Agriculture, CS, Legal, MusiQue datasets.
  • Figure 5: Knowledge base extraction process. Propositions and entities are extracted from input passages and populate the graph. Entity embeddings (used for synonym resolution) are omitted for clarity reasons.
  • ...and 5 more figures