More about HKUST
Context-Constrained Retrieval for Evidence-Grounded Retrieval-Augmented Generation
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
MPhil Thesis Defence
Title: "Context-Constrained Retrieval for Evidence-Grounded
Retrieval-Augmented Generation"
By
Mr. Kwun Hang LAU
Abstract:
Retrieval-Augmented Generation (RAG) grounds large language models with
external evidence, but reliable generation requires retrieval beyond semantic
similarity. In multi-hop and temporal questions, a retriever must recover
complete evidence chains and satisfy query-specific constraints. Existing
dense retrievers and graph-based RAG systems often rely on static similarity
scores or fixed graph transitions, causing partial retrieval, semantic drift,
and hub-node bias.
This thesis studies context-constrained retrieval for evidence-grounded RAG.
The main contribution is CatRAG, a query-adaptive graph retrieval framework
that addresses the Static Graph Fallacy in graph-based RAG. CatRAG converts a
fixed retrieval graph into a query-adaptive navigation structure through
Symbolic Anchoring, Dynamic Edge Weighting, and Key-Fact Enhancement,
allowing Personalized PageRank traversal to focus on query-relevant evidence
paths. To support the broader thesis view, we also study Time-Aware RAG for
temporal coverage and Interval-Predicate Approximate Nearest Neighbor Search
with a Unified Dominance Graph for retrieval under interval predicates.
Experiments on multi-hop question answering and fact verification show that
CatRAG improves retrieval quality, answer quality, and reasoning
completeness, measured by Full Chain Retrieval and Joint Success Rate.
Additional analysis confirms reduced hub bias and the effectiveness of the
proposed modules. Overall, this thesis shows that reliable RAG should treat
retrieval as constrained evidence selection rather than semantic
nearest-neighbor search alone.
Date: Thursday, 30 July 2026
Time: 2:00pm - 4:00pm
Venue: Room 4475
Lifts 25-26
Chairman: Prof. Raymond WONG
Committee Members: Prof. Xiaofang ZHOU (Supervisor)
Prof. Song GUO