Publication Date
Spring 2026
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Mike Wu
Second Advisor
Amith Kamath Belman
Third Advisor
Navrati Saxena
Keywords
GraphRAG, KnowledgeGraphs, LargeLanguageModels, Multi-Hop Question Answering, Query Decomposition, Retrieval-Augmented Generatio
Abstract
Multi-hop questions requiring evidence across multiple documents challenge standard Retrieval-Augmented Generation (RAG), as single vector retrieval misses cross-document connections. This research investigates whether multi-signal graph enhanced retrieval can improve multi-hop QA accuracy. It presents LOOM-RAG (Layered Ontology-Orchestrated Multi-Signal Retrieval-Augmented Generation), a graph-enhanced RAG system evaluated on the MultiHop-RAG benchmark. LOOM- RAG builds a knowledge graph with entities, typed relations, and atomic propositions, then decomposes each query into sub-questions and gathers evidence from four retrieval signals. These signals are fused via weighted combination and Reciprocal Rank Fusion, achieving 87.1% QA accuracy, a 26% improvement over Baseline RAG. An ablation study showed the multi-signal retrieval pipeline contributed 48% of the improvement, while the graph-enriched prompt contributed 52%.
Recommended Citation
Fonseca, Aldridge, "LOOM-RAG: Layered Ontology-Orchestrated Multi-Signal Retrieval-Augmented Generation for Reducing Hallucinations in LLMs" (2026). Master's Projects. 1782.
DOI: https://doi.org/10.31979/etd.hn7u-hp8x
https://scholarworks.sjsu.edu/etd_projects/1782
