How Deep GraphRAG Solves Retrieval’s Three‑Way Dilemma with Hierarchical Search

Background

Standard Retrieval‑Augmented Generation (RAG) uses dense vector retrieval, which fails on queries that require reasoning across multiple knowledge points. GraphRAG introduces a knowledge graph but creates three challenges:

Global view provides completeness but is coarse, causing detailed information to be lost in summaries.

Local view is fine‑grained but cannot traverse many hops, breaking multi‑hop reasoning.

Retrieval paths explode, leading to prohibitive latency.

Deep GraphRAG Core Techniques

Hierarchical Global→Local Retrieval : Retrieve first at the community level, then sub‑graph, then entity, pruning at each layer.

Beam‑Search Dynamic Reordering : Keep only the top‑k candidates at every step to reduce computation.

DW‑GRPO Reinforcement Learning : Treat reward weights as trainable policy parameters that are updated online, enabling a 1.5B model to match the performance of a 70B model.

Graph Construction Pipeline

Text Chunking : Sliding window of 600 tokens with 100‑token overlap.

Entity‑Relation Extraction : Use Qwen2.5‑72B‑Instruct with temperature 0 for deterministic output.

Entity Disambiguation : Compute similarity with bge‑m3; keep pairs with similarity > 0.95 and verify with an LLM.

Hierarchical Community Detection : Apply weighted Louvain recursive clustering (γ = 1.0) to produce a three‑level tree (L0, L1, L2).

Three‑Stage Retrieval Process (Algorithm 1)

Stage ① Top Community (k = 3): Input = query q + community embedding; output = top‑3 communities; similarity = cos(q, D(c)).

Stage ② Mid Community (k = 3): Expand sub‑communities; output = top‑3 sub‑communities; similarity = cos(q, D(c')).

Stage ③ Entity Layer (m = 10): Expand entities; output = top‑10 entities; similarity = cos(q, D(v)).

All steps involve only small matrix multiplications, using roughly one‑sixth the GPU resources of the Drift Search baseline.

Dynamic Reward Weighting (DW‑GRPO)

Traditional multi‑objective RL uses fixed reward weights, causing a “seesaw” effect. DW‑GRPO treats each weight as a policy parameter and updates it each training step based on a growth factor. α: Normalized slope of the reward over the past 20 steps. T = 0.1: Smoothing factor.

Reward functions:

r_rel – cross‑encoder that penalizes irrelevant answers.

r_faith – BERTScore‑F1 that penalizes hallucinations.

r_conc – 1 - len(C)/len(K) that penalizes verbosity.

Experimental Results

Performance : A 1.5B model reaches 94% of a 72B model’s Natural Questions (NQ) performance while reducing latency by 20×.

Main Metrics

HotpotQA (GQ): Deep GraphRAG 56.25% vs. Drift Search 38.75% (+17.5 pp).

NQ (Total): Deep GraphRAG 44.69% vs. Drift Search 38.05% (+6.6 pp).

GQ queries require reasoning across ≥2 communities; Deep GraphRAG handles them without degradation.

Latency Comparison

NQ‑Local: Drift Search 1.00×, Deep GraphRAG 0.14× (‑86% latency).

NQ‑Global: Drift Search 1.00×, Deep GraphRAG 0.18× (‑81.6% latency).

https://arxiv.org/pdf/2601.11144