FastInsight: Fast and Insightful Retrieval via Fusion Operators for Graph RAG

TL;DR

FastInsight tackles latency in Graph RAG on corpus graphs by introducing a retrieval taxonomy and two fusion operators, GRanker and STeX, that interleave to exploit both topology and semantics. It formalizes two fusion operators, $\mathcal{O}_{\mathrm{gm}}$ and $\mathcal{O}_{\mathrm{vgs}}$, and demonstrates substantial improvements in $R@10$ and $nDCG@10$ across diverse corpus-graph datasets with notable efficiency gains. The results show a Pareto improvement over interleaving methods, achieving up to $42$–$58\%$ reductions in query processing time while improving $R@10$ by about $11.7\%$ on challenging benchmarks, validating the approach for real-time, insightful retrieval. Overall, FastInsight advances Graph RAG by effectively integrating topology-aware and semantics-aware signals to deliver accurate, fast generation in enterprise-scale corpus graphs.

Abstract

Existing Graph RAG methods aiming for insightful retrieval on corpus graphs typically rely on time-intensive processes that interleave Large Language Model (LLM) reasoning. To enable time-efficient insightful retrieval, we propose FastInsight. We first introduce a graph retrieval taxonomy that categorizes existing methods into three fundamental operations: vector search, graph search, and model-based search. Through this taxonomy, we identify two critical limitations in current approaches: the topology-blindness of model-based search and the semantics-blindness of graph search. FastInsight overcomes these limitations by interleaving two novel fusion operators: the Graph-based Reranker (GRanker), which functions as a graph model-based search, and Semantic-Topological eXpansion (STeX), which operates as a vector-graph search. Extensive experiments on broad retrieval and generation datasets demonstrate that FastInsight significantly improves both retrieval accuracy and generation quality compared to state-of-the-art baselines, achieving a substantial Pareto improvement in the trade-off between effectiveness and efficiency.

Figures (7)

• Figure 1: Conceptual comparison of graph retrieval workflows based on retrieval operations. (a) illustrates the inputs for graph retrieval: $q$, $\mathcal{G}$, $\mathbf{v}_q$ and $\mathcal{V}$. (b)--(d) depict representative graph retrieval methods, while (e) presents our FastInsight method.
• Figure 2: Example of synthetic query generation. Red indicates textual reference to the BRAT node, while blue indicates the answer.
• Figure 3: Correlation between R@10 and Win Rate. FastInsight is the self-reference baseline (50% win rate). Dashed lines and grey areas denote linear regression fits and 95% CIs.
• Figure 4: Scatter plots illustrating the trade-off between Average QPT and R@10 on (a,c) ACL-OCL, and (b,d) SciFact-G.
• Figure 5: Impact of Topological Recall (TR) on Retrieval Performance (R@10).

Theorems & Definitions (13)

• Definition 1: Corpus Graph
• Definition 2: Vector Search, $\mathcal{O}_{\mathrm{vs}}$
• Example 1: Dense Vector Search
• Definition 3: Graph Search, $\mathcal{O}_{\mathrm{gs}}$
• Example 2: PageRank Retrieval
• Definition 4: Model-based Search, $\mathcal{O}_{\mathrm{m}}$
• Example 3: Retrieve-then-Rerank Pipeline
• Definition 5: Graph Model-based Search, $\mathcal{O}_{\mathrm{gm}}$
• Definition 6: Vector-Graph Search, $\mathcal{O}_{\mathrm{vgs}}$
• Definition 7: Topological Recall