Conversational Intent-Driven GraphRAG: Enhancing Multi-Turn Dialogue Systems through Adaptive Dual-Retrieval of Flow Patterns and Context Semantics
Ziqi Zhu, Tao Hu, Honglong Zhang, Dan Yang · 2025
Read Conversational Intent-Driven GraphRAG: Enhancing Multi-Turn Dialogue Systems through Adaptive Dual-Retrieval of Flow Patterns and Context Semantics by Ziqi Zhu, Tao Hu, Honglong Zhang, Dan Yang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-19385.txt.
Introduction
Ziqi Zhu, Tao Hu, and co-authors present CID-GraphRAG for multi-turn chat systems in this 2025 paper. Service chatbots must stay on topic across many turns while moving toward a goal. Standard retrieval systems match by meaning alone or use fixed knowledge maps, and both miss how intent shifts across a conversation.
Their system builds intent transition graphs from past dialogues that reached their goal. At reply time it retrieves along two paths: graph walks that follow likely intent flow, and meaning search over context. In tests on real service dialogues, the combined method beat meaning-only and intent-only baselines on standard scores, with a large gain in judged response quality. Neither path alone matched the pair.
Catalog record
| CREATOR | Ziqi Zhu, Tao Hu, Honglong Zhang, Dan Yang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-19385
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Conversational Intent-Driven GraphRAG: Enhancing Multi-Turn Dialogue Systems through Adaptive Dual-Retrieval of Flow Patterns and Context Semantics Ziqi Zhu, Tao Hu, Honglong Zhang, Dan Yang Abstract We present CID-GraphRAG (Conversational Intent-Driven Graph Retrieval Augmented Generation), a novel framework that addresses the limitations of existing dialogue systems in maintaining both contextual coherence and goal-oriented progression in multi-turn customer service conversations. Unlike traditional RAG systems that rely solely on semantic similarity (Conversation RAG) or standard knowledge graphs (GraphRAG), CID-GraphRAG constructs dynamic intent transition graphs from goal achieved historical dialogues and implements a dual-retrieval mechanism that adaptively balances intent-based graph traversal with semantic search. This approach enables the system to simultaneously leverage both conversional intent flow patterns and contextual semantics, significantly improving retrieval quality and response quality. In extensive experiments on real-world customer service dialogues, we employ both automatic metrics and LLM-as-judge assessments, demonstrating that CID-GraphRAG significantly outperforms both semantic-based Conversation RAG and intent-based GraphRAG baselines across all evaluation criteria. Quantitatively, CID-GraphRAG demonstrates substantial improvements over Conversation RAG across automatic metrics, with relative gains of 11 % 11\% in BLEU, 5 % 5\% in ROUGE-L, 6 % 6\% in METEOR, and most notably, a 58 % 58\% improvement in response quality according to LLM-as-judge evaluations. These results demonstrate that the integration of intent transition structures with semantic retrieval creates a synergistic effect that neither approach achieves independently, establishing CID-GraphRAG as an effective framework for addressing the challenges of maintaining contextual coherence and goal-oriented progression in knowledge-intensive multi-turn dialogues. 1 Introduction Large Language Models (LLMs) excel in text understanding and generation but struggle with domain-specific multi-turn dialogues, often producing inconsistent or contextually inappropriate responses Lewis et al. 2020 . While Retrieval-Augmented Generation (RAG) frameworks address these issues using external knowledge, conventional RAG approaches face unresolved challenges in specialized domains like customer service. Multi-turn customer service dialogues present two key challenges unaddressed by existing RAG methods. First, temporal dependencies creates complex reference chains requiring resolution across multiple turns (e.g., ”How long will that take?” needing prior context). Second, conversational objective evolves dynamically across conversation phases (e.g., information-gathering → diagnosis → solution implementation). These challenges demand a retrieval approach that captures both semantic context and conversation structure. Current research has pursued two separate approaches to address these challenges: (1) Conversation RAG methodologies[ ResearchGate 2025 , Ye et al. 2024 ] focus on optimizing dialogue history representation to enhance retrieval precision through semantic matching; (2) GraphRAG approaches[ Zhu et al. 2025 , Edge et al. 2024 , Zhang et al. 2025a ] leverage knowledge graphs to model static domain knowledge and facilitate complex reasoning. However, these methodologies have developed in parallel trajectories with fundamental limitations: Conversation RAG captures semantic similarity but lacks structural guidance for conversation flow, while GraphRAG models domain facts but not the dynamic intent transitions essential for coherent multi-turn dialogues. To address these limitations, we propose CID-GraphRAG (Conversational Intent-Driven Graph Retrieval Augmented Generation), a novel framework that integrates intent transition graphs with semantic similarity retrieval. Our approach leverages goal achieved historical conversations to model conversation dynamics through both semantic content and conversation flow, enabling more coherent and goal-oriented responses. The key contributions of our work include: • Intent Transition Graph Construction: We introduce a method to automatically build dynamic intent transition graphs from goal achieved historical dialogues, systematically cataloging hierarchical intents from both system and user utterances to create a network of effective conversational trajectories. • Dual-Pathway Adaptive Retrieval: We propose a novel retrieval approach that balances intent transition graph traversal with semantic similarity search, through an adaptive weighting mechanism. This integration enables the system to simultaneously leverage conversation structure and contextual semantics, addressing the limitations of single-signal approaches in complex multi-turn dialogues. Experiments on real-world customer service dialogues demonstrate that CID-GraphRAG significantly outperforms both semantic-only retrieval and intent-only graph approaches across all evaluation metrics, highlighting the effectiveness of our integrated dual-pathway approach. 2 Related Works 2.1 Retrieval-Augmented Generation (RAG) RAG enhances Large Language Models (LLMs) by integrating external knowledge sources, improving response accuracy and informativeness in tasks like question answering and dialogue systems Lewis et al. 2020 . Recent advancements include DH-RAG, which uses dynamic historical context to simulate human-like reasoning through long/short-term memory integration Zhang et al. 2025b , and SELF-multi-RAG, enabling LLMs to autonomously determine retrieval timing and query optimization based on conversation context Asai et al. 2023 . The MTRAG benchmark provides specialized evaluation metrics for multi-turn dialogue RAG systems Katsis et al. 2025 . Innovative applications extend RAG’s capabilities: Xu et al. combined knowledge graphs with RAG to maintain structural relationships in customer service data Xu et al. 2024 , while Veturi et al. developed a decision support system integrating real-time problem detection with FAQ retrieval Agrawal et al. 2024 . 2.2 GraphRAG GraphRAG enhances traditional RAG by integrating knowledge graphs to enable structured semantic retrieval and multi-hop reasoning, using entity-relationship representations to improve complex query handling Zhu et al. 2025 .Compared to text-based RAG, it better captures contextual dependencies in multi-turn dialogues and resolves challenges like cross-document information linking Edge et al. 2024 . Key advantages in customer service include: Enhanced retrieval accuracy through preserved structural relationships in knowledge graphs Xu et al. 2024 ; Effective processing of multi-hop queries requiring cross-fragment reasoning [ Zhu et al. 2025 , Zhang et al. 2025a ]; Context continuity maintenance across dialogue turns via entity relationship tracking Xu et al. 2024 , complementing DH-RAG’s dynamic context management Zhang et al. 2025b . The framework mitigates text segmentation issues in traditional RAG while improving answer interpretability through visualized reasoning paths Linders and Tomczak 2025 , particularly benefiting complex service scenarios with interlinked conversational dependencies. 2.3 Conversation RAG Conversation RAG addresses multi-turn dialogue challenges in customer service, where user queries often depend on historical interactions, requiring systems to manage evolving contexts and avoid redundant retrievals ResearchGate 2025 . These techniques enhance contextual relevance and response coherence across dialogue turns Ye et al. 2024 . Key advancements include: • Context-aware frameworks: DH-RAG Zhang et al. 2025b and SELF-multi-RAG Asai et al. 2023 optimize historical context modeling; • Structured retrieval: KG-RAG Xu et al. 2024 leverages knowledge graphs for improved accuracy; • Query decomposition: Collab-RA Xu et al. 2025 breaks down complex questions via collaborative training; • Performance benchmarks: RAD-Bench Kuo et al. 2024 and MTRAG Katsis et al. 2025 enable systematic evaluation. ChatQA Liu et al. 2024 demonstrates open-source models can achieve state-of-the-art conversational QA through retriever fine-tuning. While standard RAG remains foundational, GraphRAG excels in structured reasoning and Conversation RAG advances context-aware interactions, yet no universal solution exists. 3 Methodology CID-GraphRAG innovatively integrates Conversation RAG and GraphRAG to address the challenge in current multi-turn customer service dialogue systems of simultaneously maintaining contextual coherence and goal-oriented progression. Unlike traditional RAG approaches, CID-GraphRAG offers the following key advantages: 1. IntenGraph: Differing from conventional graphs, we propose a novel dual-layer intent graph. This replaces primitive nodes with dual-layer intent nodes, dynamically extracting expert intent transition processes from high-quality historical assistant dialogues to generate a domain-specific expert intent transition map. 2. Dual-layer Retrieval: Moving beyond traditional RAG systems that rely solely on semantic similarity (Conversation RAG) or static knowledge graphs (GraphRAG), we innovatively integrate the intent transition map with semantic similarity retrieval. This dual-layer retrieval mechanism enables mutually-assisted,