How CID-GraphRAG Boosts Multi‑Turn AI Customer Service with Dual‑Layer Retrieval

Background

In the wave of digital transformation, intelligent customer service and outbound calling have become core tools for enterprises to reduce costs, increase efficiency, and improve customer experience. They use multi‑turn dialogue technology to simulate human interaction, providing 24/7 instant Q&A, business handling, and proactive customer outreach.

CID-GraphRAG Framework

The proposed CID-GraphRAG (Conversation‑Intent‑Driven Graph Retrieval‑Augmented Generation) merges intent‑based graph matching with semantic search, constructing a dual‑path retrieval architecture that addresses the difficulty of maintaining context coherence and goal‑driven progression in multi‑turn dialogues.

Core Technical Pipeline

Speech Input & Transcription The system first uses ASR to convert customer speech into accurate text in real time.

Semantic Understanding & Enhancement Specialized terms are normalized via hot‑word replacement to ensure precise recognition of key information.

Large‑Model‑Based Response Generation LLMs retrieve relevant information from a proprietary knowledge base using RAG techniques and generate highly relevant replies.

Speech Output The generated text is synthesized into natural, fluent speech via streaming TTS for the final customer interaction.

Current systems still suffer from two main issues: (1) replies only address the immediate user query, ignoring broader context, leading to off‑topic answers; (2) repetitive responses degrade the human‑machine experience. Consequently, LLM‑based agents fall short of expert‑level understanding, multi‑turn guidance, and human‑like interaction.

Intent Graph Construction

Expert multi‑turn dialogues are processed with LLMs to extract primary and secondary intents, forming a two‑layer intent graph where primary intents define the overall conversation direction and secondary intents capture fine‑grained details. This dual‑layer design prevents semantic conflicts and enables precise intent classification.

Dual‑Layer Retrieval Mechanism

The system performs two complementary retrievals:

Intent‑Based Retrieval Using the identified intents, the Intent Graph is queried to find the next most probable agent intent, providing strategic, goal‑oriented guidance.

Semantic Retrieval Vector similarity matches the current dialogue context against a large knowledge base, ensuring responses stay contextually relevant.

Results from both paths are weighted (parameter α) to produce a combined score, selecting top‑k historical dialogues as few‑shot examples for the final LLM response generation.

Response Generation

The final prompt incorporates the user question, dialogue memory, detected user intent, few‑shot examples from the dual retrieval, and explicit instructions, enabling the LLM to produce replies that are natural, expert‑like, and strategically driven toward successful task completion.

Evaluation

CID-GraphRAG was benchmarked against Direct LLM, Intent RAG, and Conversation RAG on 126 test cases. Both LLM preference assessments and automatic metrics (BLEU, ROUGE, METEOR, BERTScore) show CID-GraphRAG leading in retrieval quality (73 wins vs. 51) and response quality (60 wins vs. 38). The dual‑layer approach consistently outperforms single‑path baselines, confirming the synergistic benefit of combining intent graphs with semantic similarity.

Conclusion & Future Work

CID-GraphRAG effectively bridges the gap between context‑aware dialogue and goal‑driven assistance, delivering superior answer relevance and user experience. Future directions include integrating reinforcement learning to dynamically optimize intent transition probabilities and extending the framework to broader, cross‑domain conversational scenarios.