11: Graph RAG – Retrieval Augmented Generation (CS6101 WING.NUS)

Описание: (Lecture Starts at xx:xx)

Slides for this session: http://soc-n.us/cs6101-t2510-w11
Video Playlist: https://soc-n.us/cs6101-t2510-playlist

CS6101/DYC1401 Retrieval-Augmented Generation
Week 11: 30 Oct 2025 AY 25/26 Sem 1 (T2510)

Summary by Zoom AI
Enhancing Model Performance Strategies
The team discussed various aspects of their machine learning model, including threshold content, granularity mismatch, and fine-tuning techniques. They addressed issues with low-confidence scores and overfitting, and explored methods for improving model performance across different modalities. The conversation touched on topics such as contrastive loss and cross-encoder models, with a focus on enhancing the matching of words in queries to multimodal items.

Graph Rank and Course Updates
The meeting covered updates on a course project, including discussions on hours spent, upcoming deadlines, and the STEPS event. Kan announced that external guests can attend the STEPS event by registering online. The class reviewed the progress of various team projects and introduced a guest speaker, Alice Galatros, who discussed graph rank and its applications in information retrieval. Andrew Neil presented on graph rank, introducing concepts like knowledge graphs and trust metrics, and the class engaged in an interactive discussion on the benefits of graph rank over traditional rank methods.

GraphRank and LightGraph Overview
Kan discussed the concept of GraphRank, a method for modeling interconnected information as nodes and relationships in a graph, which is particularly useful for handling multi-hop questions and complex queries. He explained the limitations of traditional rank systems and highlighted the advantages of GraphRank, such as its ability to capture high-level overviews and find connections between entities. Kan also introduced the framework of GraphRank, which includes a query processor, retriever, and a language model for generating answers. He critiqued the evaluation method used in the GraphRank paper, noting that it focused on sense-making rather than multi-hop queries. Kan concluded by introducing LightGraph, an improved version of GraphRank that uses an LM for entity extraction and employs a high-level and low-level query approach to retrieve information efficiently.

Graph Indexing and Knowledge Retrieval
Kan explained the process of indexing and creating knowledge graphs, focusing on the Leiden algorithm for community detection in large graphs. Li discussed the graph data source component of the Graph IG framework, highlighting its advantages over traditional methods and various representation formats. They covered examples of knowledge graphs and scientific graphs, emphasizing their importance in enabling modular structure retrieval and reducing harmonization. Li concluded by mentioning future directions for graph data source development, including better construction tools and updates.

Graph Algorithms for Query Processing
The meeting focused on graph algorithms and query processing in the context of knowledge representation and retrieval. Kan explained the importance of graph quality and the need for good data to ensure the effectiveness of graph algorithms. They discussed the process of converting natural language questions into a format suitable for graph processing, outlining five sub-processes: entity recognition, query structuration, query relation attraction, query decomposition, and query expansion.

Graph Retrieval Methods Overview
Van-Hoang presented on the retriever module of the graph rack, discussing both heuristic-based and learning-based retrieval methods. He covered entity linking, relational matching, and graph traversal, as well as a KGBT framework for reasoning on knowledge graphs using LLMs. The presentation also included a discussion of GNN-based retrieval methods and their combination with other approaches to improve context for LLMs.

Knowledge Graph Organizers and Generators
Luis presented on knowledge graph organizers and generators, covering four types of organizers: pruning, re-ranking, augmenting, and verbalizing. He explained various pruning methods including semantic, structural, syntactic, and dynamic approaches, as well as re-ranking techniques using pre-trained models. For augmentation, he discussed both feature and structural augmentation approaches, including examples from PIPNet and UniAug. In the generator section, he covered LLM-based, discrimination-based, and graph-based approaches, explaining how to integrate knowledge graphs with text embeddings and other architectures.

00:00:00 Kahoot! from W10
00:20:00 Section 1: Introduction to GraphRAG (Indraneel)
00:40:00 Section 2: Graph Data Source (Li Zizhen)
01:00:00 Section 3: Query Processor (Lin Hong Yi)
01:20:00 Section 4: GraphRAG Retriever (Van-Hoang Nguyen)
01:40:00 Section 5: Organizer and Generator (Luis Frentzen Salim)