Introduction
Retrieval-Augmented Generation (RAG) has transitioned from simple vector-based retrieval (NaiveRAG) to sophisticated graph-based structures like GraphRAG. While these advancements improve the capture of entity relationships, they often falter in Query-Focused Summarization (QFS). Current systems frequently suffer from two primary issues: insufficient context due to fragmented entity retrieval and query-unaware summarization that introduces irrelevant "noise" from pre-generated community summaries. This paper introduces FG-RAG, a framework designed to overcome these hurdles by providing context-aware retrieval and fine-grained, query-driven synthesis.
Methodology: The FG-RAG Framework
FG-RAG (Fine-Grained Graph RAG) addresses the limitations of existing graph-based RAG systems through a two-stage process: Context-Aware Entity Expansion and Query-Level Fine-Grained Summarization.
Context-Aware Entity Expansion
Standard retrieval often misses the broader context necessary for a comprehensive answer. FG-RAG identifies "weak-context" entities (direct query matches) and employs a Breadth-First Search (BFS) to retrieve "strong-context" neighbors within the knowledge graph. This ensures the model understands the background and relationships surrounding the core facts, providing the necessary depth for complex queries.
Query-Level Fine-Grained Summarization
To avoid the noise of general summaries, FG-RAG utilizes an agentic pipeline that decomposes the user's query into entity-specific sub-problems:
- Entity Extraction: Identifying key entities in the user query.
- Question Formulation: Generating specific sub-questions for each entity (e.g., "What are the marketing channels for [Honey]?").
- Fine-Grained Synthesis: Retrieving descriptions for these sub-questions to build targeted mini-summaries.
- Final Answer Generation: Synthesizing these relevant summaries into a final, cohesive response.
Experimental Evaluation
The framework was evaluated using the UltraDomain benchmark (covering Agriculture, Art, Legal, and Mixed domains) and a News dataset for multi-hop reasoning. Performance was measured via a "win-rate" approach using GPT-4o-mini as a judge across four dimensions: Comprehensiveness, Diversity, Empowerment, and Overall quality.
Key Results and Insights
FG-RAG consistently outperformed state-of-the-art baselines like LightRAG and Microsoft’s GraphRAG, often achieving win rates between 60% and 80% across various LLMs.
Small Model Empowerment: A significant finding is FG-RAG's ability to boost smaller LLMs (e.g., Qwen-1.5B and 3B). By providing highly relevant, noise-free context, it reduces the reasoning burden, allowing smaller models to achieve performance comparable to much larger counterparts.
Efficiency: Despite its multi-step pipeline, FG-RAG is highly token-efficient. In tests on the "Mix" dataset, it utilized over 50% fewer query tokens than LightRAG while maintaining superior answer quality. This efficiency makes it a viable candidate for production environments where cost and latency are critical.
Conclusion
FG-RAG represents a significant step forward in making graph-based RAG more precise and efficient. By moving away from static community summaries toward dynamic, query-driven synthesis, it provides more structured and nuanced answers. While the multi-step process introduces potential latency, the gains in accuracy and token efficiency offer a compelling trade-off. Future work will focus on optimizing the question formulation process and exploring more specialized agentic structures.
Summary updated