Advanced Retrieval-Augmented Generation in Multi-Turn Dialogues and Heterogeneous Multi-Cloud Computing Environments

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "Advanced Retrieval-Augmented Generation in Multi-Turn Dialogues and 
Heterogeneous Multi-Cloud Computing Environments"

By

Miss Feiyuan ZHANG


Abstract:

Traditional Retrieval-Augmented Generation (RAG) systems face significant 
limitations in multi-turn dialogue scenarios, failing to effectively capture 
and utilize dynamic conversational context. This thesis addresses both 
algorithmic innovations and distributed deployment challenges for advanced 
conversational AI systems.

We introduce DH-RAG (Dynamic Historical Context-Powered Retrieval-Augmented 
Generation), a novel method specifically designed for multi-turn dialogues 
that integrates static external knowledge with dynamic conversational 
context. DH-RAG comprises a History-Learning based Query Reconstruction 
Module and a Dynamic History Information Updating Module, supported by a 
Dynamic Historical Information Database employing Historical Query 
Clustering, Hierarchical Matching, and Chain of Thought Tracking strategies.

Experimental evaluation demonstrates that DH-RAG significantly outperforms 
conventional methods across multiple benchmarks, achieving improvements 
ranging from 25% to 85% in response quality metrics. On domain-specific 
datasets, DH-RAG achieves BLEU scores of 4.10 and F1 scores of 27.83, 
representing 215.38% and 58.13% improvements over baseline methods.

Beyond algorithmic contributions, we design and implement a comprehensive 
distributed architecture for heterogeneous multi-cloud environments. Our 
dual- microservice architecture enables independent scaling while Our 
intelligent GPU scheduling system achieves 22% efficiency improvement and 
18% cost optimization. The three-node deployment spanning Hong Kong, 
Shandong, and Jiangsu demonstrates 210% throughput improvement compared to 
single-machine deployment.

This integrated approach validates that sophisticated conversational AI 
algorithms can be effectively deployed in distributed environments while 
maintaining algorithmic integrity and practical scalability, establishing a 
research paradigm bridging theoretical advances with practical engineering 
solutions. and other decentralized applications.


Date:                   Friday, 1 August 2025

Time:                   4:00pm - 6:00pm

Venue:                  Room 3494
                        Lifts 25/26

Chairman:               Prof. Song GUO

Committee Members:      Prof. Kai CHEN (Supervisor)
                        Prof. Gary CHAN