Youtu-GraphRAG: Unified Graph Augmented Framework

• Youtu-GraphRAG is a unified framework for graph retrieval-augmented generation that integrates schema-driven knowledge extraction, community detection, and agentic retrieval.
• It employs dual-perception community detection and iterative agent-based reasoning to enable robust multi-hop reasoning and seamless domain transfer with minimal manual schema intervention.
• The system demonstrates significant performance gains, achieving up to 90.71% token cost savings and a 16.62% improvement in top-k accuracy over previous GraphRAG methods.

Youtu-GraphRAG is a vertically unified agentic framework for graph retrieval-augmented generation, designed to enhance LLMs in complex reasoning tasks by integrating schema-guided knowledge extraction, dual-perception community detection, schema-driven agentic retrieval, and rigorous anonymization for knowledge leakage assessment. The system targets robust multi-hop reasoning and domain transfer with minimal schema intervention, achieving marked improvements in both efficiency and accuracy relative to prior GraphRAG methods (Dong et al., 27 Aug 2025).

1. Seed Graph Schema for Targeted Knowledge Extraction

At the foundation of Youtu-GraphRAG is the explicit definition of a compact seed schema, formalized as a triple:

$$\mathcal{S} \triangleq \langle \mathcal{S}_e, \mathcal{S}_r, \mathcal{S}_{attr}\rangle$$

where $\mathcal{S}_e$ defines the entity types (e.g., Person, Disease), $\mathcal{S}_r$ the relation types (e.g., treats, causes), and $\mathcal{S}_{attr}$ the attribute types (e.g., occupation, gender). Knowledge extraction operates under constraints dictated by the Cartesian product $$\mathcal{S}_e \times \mathcal{S}_r \times \mathcal{S}_{attr}$$, ensuring that only schema-compliant triples are constructed:

$$\mathcal{T}(d) = \{ (h, r, t), (e, r_{attr}, e_{attr}) \mid f(h), f(t), f(e) \in \mathcal{S}_e, r, r_{attr} \in \mathcal{S}_r, e_{attr} \in \mathcal{S}_{attr} \}$$

Scalability and adaptability are addressed via automatic schema expansion. Upon encountering new relational patterns in domain-shifting data, candidate schema updates $\Delta \mathcal{S}$ are conditionally accepted when an LLM-based extraction agent assigns sufficient confidence (threshold $\mu$) to ensure schema extension remains targeted and avoids spurious expansion.

2. Dual-Perception Community Detection and Hierarchical Knowledge Tree Construction

Youtu-GraphRAG employs a novel community detection algorithm that fuses structural topology with subgraph semantics, yielding a hierarchical organization of knowledge for both top–down filtering and bottom–up reasoning:

1. Entity Representation: Each entity $e_i$ aggregates information from its one-hop neighborhood $\mathcal{N}_i$:

$$e_i = \frac{1}{|\mathcal{N}_i|}\sum_{(e_i, r, e_j) \in \mathcal{N}_i} [ e_i \| r_{ij} \| e_j ]$$

1. Initial Clustering: K-means clusters are formed adaptively:

$$k = \min( \max(2, \lfloor |\mathcal{E}|/\beta \rfloor ), \eta )$$

1. Dual-Perception Scoring: The affinity of node $e_i$ for community $\mathcal{C}_m$ is:

$$\phi(e_i, \mathcal{C}_m) = \mathbb{S}_r(e_i, \mathcal{C}_m) \oplus \lambda \mathbb{S}_s(e_i, \mathcal{C}_m)$$

where $\mathbb{S}_r(\cdot,\cdot)$ is the Jaccard similarity over incident relations and $\mathbb{S}_s(\cdot,\cdot)$ is the cosine similarity of transformed semantic embeddings. Centers are updated to nodes with maximum $\phi$, and merging proceeds if mean score differences are below $\epsilon$.

This produces a multi-level knowledge tree (Community, Keywords, Triples, Attributes) that preserves both local and global knowledge structure and supports efficient retrieval for LLM reasoning.

3. Agentic Retrieval: Schema-Driven Decomposition, Iteration, and Reflection

The retrieval process in Youtu-GraphRAG is orchestrated by an agentic retriever that interprets the graph schema and orchestrates query decomposition alongside iterative reasoning:

• Schema-Enhanced Query Decomposition: The retriever decomposes complex input queries $q$ into a set of atomic sub-queries $\{q_1, q_2, ..., q_i\}$ aligned with schema components.
• Iterative Reasoning with Memory: The agent, defined as $$\mathcal{A} = \langle\mathcal{S}, \mathcal{H}, f_{LLM}\rangle$$, maintains a history $\mathcal{H}$ of previous reasoning steps and retrieval results, updating its state at iteration $t$ via

$$\mathcal{A}^{(t)} = f_{LLM}(q^{(t)}_{Reasoning}, \mathcal{H}^{(t-1)})$$

• Parallel Retrieval Strategies: Includes entity matching (cosine similarity maximization), triple matching, community filtering, and constrained depth-first graph traversal, supporting both direct evidence lookup and long-range multi-hop reasoning.

This agentic retriever supports advanced reasoning via iterative cycles of sub-query execution and reflection upon intermediate results.

4. Anonymity Reversion and Knowledge Leakage Assessment

To rigorously evaluate retrieval-augmented reasoning and prevent LLMs from leveraging parametric knowledge (“knowledge leakage”), Youtu-GraphRAG introduces an “Anonymity Reversion” protocol:

• Entity references (e.g., person or location names) in gold-standard queries and evidence are replaced with anonymized tags ([PERSON#277], [LOCATION#759]).
• During inference, LLMs must resolve anonymized forms to their true identities using only context from the retrieved graph and supporting evidence, not pre-trained parametric memory.

Evaluations are conducted against the AnonyRAG-CHS and AnonyRAG-ENG datasets, ensuring that performance metrics reflect genuine retrieval and reasoning rather than memorization.

5. Efficiency and Performance Metrics

Empirical assessments on benchmarks including HotpotQA, 2WikiMultiHopQA, MuSiQue, G-Bench, and AnonyRAG variants show that Youtu-GraphRAG achieves substantial advances:

Metric	Youtu-GraphRAG	Best Baseline
Token Cost Saving	up to 90.71%	–
Top-k Accuracy Gain	+16.62%	–

• Token Efficiency: Youtu-GraphRAG reduces token consumption by up to an order of magnitude in graph extraction and community detection phases.
• Pareto Improvement: Demonstrated by Pareto-frontier plots, the approach simultaneously achieves lower resource use and higher QA accuracy relative to prior GraphRAG approaches.
• Domain Adaptability: The vertically unified pipeline allows schema expansion and community reconfiguration with minimal manual intervention, supporting seamless transfer to new domains.

6. Integration and Theoretical Significance

Youtu-GraphRAG unifies extraction, organization, and retrieval under a schema-aware, agentic pipeline to address the often disjointed or narrowly optimized stages of prior GraphRAG systems. The dual-perception clustering improves community structure by leveraging both topology and semantic signals, critical for pathways in multi-hop reasoning. Agentic retrieval ensures that the reasoning chain is both schema-constrained and capable of reflection and iteration, supporting explainable and robust inference. The Anonymity Reversion protocol and dedicated datasets provide a practical methodology for detecting and eliminating the confounding influence of model memorization on retrieval-augmented reasoning benchmarks.

The framework's adaptability across six challenging benchmarks, marked by a consistent improvement of 16.62% in top-k accuracy and up to 90.71% token cost saving, positions it as an effective paradigm for scalable, efficient, and transferable complex reasoning in retrieval-augmented LLM systems.