Deep GraphRAG: A Balanced Approach to Hierarchical Retrieval and Adaptive Integration

Graph-based Retrieval-Augmented Generation (GraphRAG) frameworks face a trade-off between the comprehensiveness of global search and the efficiency of local search. Existing methods are often challenged by navigating large-scale hierarchical graphs, optimizing retrieval paths, and balancing exploration-exploitation dynamics, frequently lacking robust multi-stage re-ranking. To overcome these deficits, we propose Deep GraphRAG, a framework designed for a balanced approach to hierarchical retrieval and adaptive integration. It introduces a hierarchical global-to-local retrieval strategy that integrates macroscopic inter-community and microscopic intra-community contextual relations. This strategy employs a three-stage process: (1) inter-community filtering, which prunes the search space using local context; (2) community-level refinement, which prioritizes relevant subgraphs via entity-interaction analysis; and (3) entity-level fine-grained search within target communities. A beam search-optimized dynamic re-ranking module guides this process, continuously filtering candidates to balance efficiency and global comprehensiveness. Deep GraphRAG also features a Knowledge Integration Module leveraging a compact LLM, trained with Dynamic Weighting Reward GRPO (DW-GRPO). This novel reinforcement learning approach dynamically adjusts reward weights to balance three key objectives: relevance, faithfulness, and conciseness. This training enables compact models (1.5B) to approach the performance of large models (70B) in the integration task. Evaluations on Natural Questions and HotpotQA demonstrate that Deep GraphRAG significantly outperforms baseline graph retrieval methods in both accuracy and efficiency.

💡 Research Summary

Deep GraphRAG tackles the long‑standing trade‑off in graph‑based Retrieval‑Augmented Generation (GraphRAG) between the breadth of global search and the speed of local search. The authors introduce a hierarchical “global‑to‑local” retrieval pipeline that explicitly separates the graph into three levels of abstraction: (1) inter‑community filtering, (2) community‑level refinement, and (3) entity‑level fine‑grained search. The pipeline is driven by a beam‑search‑inspired dynamic re‑ranking module that continuously prunes candidates, ensuring that only the most promising sub‑graphs are explored at each stage.

Graph construction starts with a sliding‑window chunking of the raw corpus (600‑token windows with 100‑token overlap). Each chunk is processed by Qwen2.5‑72B‑Instruct to extract entities and directed relationships, but instead of classic RDF triples the system generates concise natural‑language edge descriptions to preserve richer semantics. Entity resolution combines cosine similarity on bge‑m3 embeddings (threshold > 0.95) with a secondary LLM discriminator to merge duplicate nodes with near‑perfect precision. The resulting base graph G is then partitioned into a three‑level community hierarchy C using the weighted Louvain algorithm (resolution γ = 1.0). Community embeddings are obtained by mean‑pooling sub‑community vectors, while each entity’s representation is a concatenation of its own embedding and its parent community embedding, yielding context‑aware vectors D_ctx(v).

The retrieval algorithm (Algorithm 1) proceeds as follows:

1. Phase 1 – Inter‑community filtering: a lightweight cross‑encoder quickly scores all top‑level communities; the top‑k (k = 3) are retained.
2. Phase 2 – Community‑level refinement: the selected communities are expanded to their sub‑communities and re‑scored with the bge‑reranker‑v2‑m3, which captures subtle semantic alignment between the query and community summaries. The top‑k sub‑communities are kept.
3. Phase 3 – Entity‑level fine‑grained search: candidate entities within the retained sub‑communities are evaluated using the concatenated context‑aware embedding D_ctx(v) and cosine similarity with the query. The top‑m entities are output.
4. Phase 4 – Knowledge integration: the selected entities are fed into a Knowledge Integration Module that produces a distilled answer C.

A key novelty is the Dynamic Weighting Reward GRPO (DW‑GRPO) reinforcement‑learning framework used in the integration module. Traditional GRPO applies static weights to multiple reward signals (relevance, faithfulness, conciseness), which often leads to a “seesaw effect” where easy‑to‑optimize rewards dominate and the model neglects harder objectives. DW‑GRPO replaces static weights with a policy‑aware adaptive weighting mechanism. For each reward r_j, the algorithm computes a sliding‑window change rate α_j(t) based on the linear‑fit slope of the reward trajectory. Rewards that exhibit slower growth receive higher softmax‑scaled weights (temperature T), thereby directing the policy to focus on under‑performing aspects. The advantage estimate is adjusted accordingly (Eq. 5), and the overall reward is a weighted sum e = ∑ w_j r_j.

Experiments on Natural Questions (NQ) and HotpotQA evaluate both accuracy (Exact Match, EM) and latency. Questions are categorized into Local (LQ), Global (GQ), and Comprehensive (CQ) to assess performance across reasoning depths. Baselines include Local Search (dense vector over all entities), Global Search (map‑reduce style summarization), and Drift Search (recursive hierarchical search).

Results show that Deep GraphRAG with the 72 B integration model achieves state‑of‑the‑art EM‑Total scores: 44.69 % on NQ and 44.67 % on HotpotQA, outperforming all baselines. The gains are most pronounced on Global Questions, where the model reaches 55.08 % (NQ) and 56.25 % (HotpotQA) EM, demonstrating its ability to aggregate information across multiple communities. On Comprehensive Questions, performance is competitive but not uniformly superior, indicating that the hierarchical summarization can sometimes obscure fine‑grained facts—a limitation the authors acknowledge.

A compact 1.5 B model trained with DW‑GRPO attains 42.36 % EM on NQ and 38.44 % EM on HotpotQA, roughly 94 % of the 72 B model’s performance, confirming that dynamic reward weighting enables lightweight models to approach large‑model quality. Latency analysis reveals an 86 % reduction compared to Local Search and an 81.6 % reduction compared to Global Search on NQ, highlighting the efficiency of the beam‑search‑guided hierarchy.

The paper concludes that the combination of (1) a rigorous graph construction and community hierarchy, (2) a three‑stage global‑to‑local retrieval with beam‑search re‑ranking, and (3) the DW‑GRPO adaptive RL scheme, provides a balanced solution to the efficiency‑accuracy dilemma in GraphRAG. Future work will explore finer preservation of local facts within comprehensive queries, online updating of community structures, and more sophisticated, possibly non‑linear, reward‑rate estimation for DW‑GRPO.