Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BubbleRAG: Evidence-Driven Retrieval-Augmented Generation for Black-Box Knowledge Graphs

Duyi Pan, Tianao Lou, Xin Li, Haoze Song, Yiwen Wu, Mengyi Deng, Mingyu Yang, Wei Wang

Abstract

Large Language Models (LLMs) exhibit hallucinations in knowledge-intensive tasks. Graph-based retrieval augmented generation (RAG) has emerged as a promising solution, yet existing approaches suffer from fundamental recall and precision limitations when operating over black-box knowledge graphs -- graphs whose schema and structure are unknown in advance. We identify three core challenges that cause recall loss (semantic instantiation uncertainty and structural path uncertainty) and precision loss (evidential comparison uncertainty). To address these challenges, we formalize the retrieval task as the Optimal Informative Subgraph Retrieval (OISR) problem -- a variant of Group Steiner Tree -- and prove it to be NP-hard and APX-hard. We propose BubbleRAG, a training-free pipeline that systematically optimizes for both recall and precision through semantic anchor grouping, heuristic bubble expansion to discover candidate evidence graphs (CEGs), composite ranking, and reasoning-aware expansion. Experiments on multi-hop QA benchmarks demonstrate that BubbleRAG achieves state-of-the-art results, outperforming strong baselines in both F1 and accuracy while remaining plug-and-play.

BubbleRAG: Evidence-Driven Retrieval-Augmented Generation for Black-Box Knowledge Graphs

Abstract

Large Language Models (LLMs) exhibit hallucinations in knowledge-intensive tasks. Graph-based retrieval augmented generation (RAG) has emerged as a promising solution, yet existing approaches suffer from fundamental recall and precision limitations when operating over black-box knowledge graphs -- graphs whose schema and structure are unknown in advance. We identify three core challenges that cause recall loss (semantic instantiation uncertainty and structural path uncertainty) and precision loss (evidential comparison uncertainty). To address these challenges, we formalize the retrieval task as the Optimal Informative Subgraph Retrieval (OISR) problem -- a variant of Group Steiner Tree -- and prove it to be NP-hard and APX-hard. We propose BubbleRAG, a training-free pipeline that systematically optimizes for both recall and precision through semantic anchor grouping, heuristic bubble expansion to discover candidate evidence graphs (CEGs), composite ranking, and reasoning-aware expansion. Experiments on multi-hop QA benchmarks demonstrate that BubbleRAG achieves state-of-the-art results, outperforming strong baselines in both F1 and accuracy while remaining plug-and-play.
Paper Structure (21 sections, 2 theorems, 8 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 21 sections, 2 theorems, 8 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Graph-Based RAG
  4. Black-Box Knowledge Graphs
  5. Motivations
  6. Problem Formulation
  7. The BubbleRAG Framework
  8. Data Preparation
  9. Semantic Anchor Grouping
  10. Candidate Evidence Graph Discovery
  11. Candidate Evidence Graph Ranking
  12. Reasoning-Aware Expansion
  13. Answer Generation
  14. Experiments
  15. Experimental Setup
  16. ...and 6 more sections

Key Result

Theorem 1

The Optimal Informative Subgraph Retrieval problem is NP-hard.

Figures (4)

  • Figure 1: Three challenges in black-box Knowledge Graph retrieval that limit recall and precision: (a) Semantic instantiation uncertainty when grounding query concepts into KG entities, risking recall loss; (b) Structural path uncertainty when determining relevant relational chains, risking recall loss; and (c) Evidential comparison uncertainty when ranking candidates based on implicit evidence, risking precision loss
  • Figure 2: Pipeline of BubbleRAG.
  • Figure 3: An example of the Candidate Evidence Graph (CEG) generated by Bubble Expansion. Different highlight colors in the query represent different extracted keywords, which are mapped to their corresponding semantic anchor groups in the graph. Elements with red borders indicate the nodes and edges that have been successfully incorporated into a CEG.
  • Figure 4: An example of Candidate Evidence Graph (CEG) Ranking. Different highlight colors in the query represent different semantic anchor groups, each assigned an importance weight. The final score of a CEG is determined by two components: its average semantic cost and a structural incompleteness penalty derived from missing concept groups. By simply adjusting the hyperparameter $\alpha$, the system can dynamically alter the ranking results to support AND operation and OR operation. The algorithm naturally supports compare queries by selecting the top-$n$ CEGs.

Theorems & Definitions (8)

  • Example 1
  • Example 2
  • definition 1: Optimal Informative Subgraph Retrieval (OISR)
  • Example 3
  • Example 4
  • Example 5
  • Theorem 1
  • Theorem 2