NemoClaw Knowledge Wiki

Discover AI channel - Graph RAG evolved

6 min read

https://www.youtube.com/watch?v=oetP9uksUwM This video provides a comprehensive overview of the evolution of Retrieval-Augmented Generation (RAG) systems, from foundational RAG to GraphRAG, LightRAG, and the latest development, PathRAG. The speaker aims to explain how these advanced RAG techniques work and provides relevant code resources. Here’s a breakdown of the key points: 1. Limitations of Traditional RAG (0:17 - 0:48)

  • Traditional RAG, despite various optimization techniques (chunking strategies, pre/post-retrieval optimizations, embedding model fine-tuning), often struggles with robustness.
  • A critical flaw in current graph-based RAG approaches is “information overload.” They retrieve excessively broad subgraphs, leading to noisy prompts, increased computational cost, and suboptimal Large Language Model (LLM) performance. This is particularly evident in global queries (e.g., “What are the main themes in this dataset?”), which traditional RAG fails to answer effectively.

2. GraphRAG (1:40 - 3:53)

  • Introduced as a graph-based approach to query-focused summarization (Microsoft Research).
  • Key Idea: It uses an LLM to build a graph index in two stages: Derive an entity knowledge graph: Extracts entities and relationships from source documents. Pre-generate community summaries: Creates summaries for all groups of closely related entities (thematic clusters).
  • When given a question, relevant community summaries are used to generate partial responses, which are then summarized into a final global answer.
  • GraphRAG shows substantial improvements over vector RAG baselines in terms of comprehensiveness and diversity of answers.

3. LightRAG (3:53 - 7:08)

  • Developed by Peking University of Posts and Telecommunications (November 2024).
  • Key Idea: It incorporates graph structures into text indexing and retrieval processes.
  • Graph-Based Text Indexing: It processes human-readable text by: Deduplicating information. LLM Profiling (e.g., identifying “A beekeeper is a person who…”). Entity & Relation Extraction (e.g., “beekeeper observes bees”).
  • This process builds an index graph used for retrieval, enriching nodes and edges with additional information like source, description, and keywords.
  • LightRAG employs a dual-level retrieval paradigm using high-level and low-level keys for queries and LLM processing.
  • The project has an active GitHub community with useful graph visualization tools.

4. The “Indexing Graph” (7:08 - 8:37)

  • The speaker emphasizes that the “indexing graph” is a specialized knowledge graph designed specifically for indexing and retrieval in RAG.
  • Unlike traditional RAG (which indexes pages/paragraphs by keywords), the indexing graph is like a “concept map.” Nodes: Represent key concepts (e.g., “photosynthesis,” “chlorophyll”). Edges: Represent relationships (e.g., “photosynthesis requires sunlight”). Textual Chunks: Crucially, each concept and relationship is linked back to the relevant sentences or paragraphs in the original text that explain them.

5. PathRAG - The Newest Evolution (9:40 - 19:37)

  • Published in February 2025 by authors from Beijing University, University of Hong Kong, and Northeastern University (some also involved in LightRAG).
  • Overarching Goal: To overcome the limitations of previous graph-based RAG methods (redundancy, flat structure of retrieved info, suboptimal logicality, and coherence).
  • Aims to be a better graph-based RAG by: Reducing Noise: Alleviating redundant and irrelevant information. Reducing Token Consumption: Retrieving and using less information more efficiently. Improving Answer Quality: Generating more logical, coherent, and higher-quality responses.
  • PathRAG’s Core Mechanics (Details): Identifies keywords in the user’s query. Searches within the indexing graph to retrieve relevant nodes based on these keywords (finding “starting points”). Applies a Flow-based Pruning Algorithm with Distance Awareness: This mechanism simulates “flow” or “resource” through the graph, starting from retrieved nodes. It’s a “pruning” algorithm that actively eliminates or filters out paths deemed less important or relevant. It’s “distance aware,” prioritizing shorter and more directly connected paths (which are likely more semantically related and less noisy). Each path is assigned a “reliability score.” For each retrieved path: It fetches the textual chunks associated with each node and edge along the path (leveraging the indexing graph’s associated text). It then sequentially concatenates these textual chunks in the order they appear in the path (node, edge, node, edge, etc.). This forms a “textual relational path” – a human-readable textual representation of the path and its associated information, making the textual information explicit for the LLM.
  • Overall Framework (3 Main Stages): Node Retrieval Stage: Relevant nodes are retrieved from the indexing graph based on query keywords, using dense vector matching and cosine similarity in the semantic embedding space. (The speaker notes this means it still relies on vector space, similar to classical RAG). Path Retrieval Stage: A flow-based pruning algorithm extracts key relational paths between each pair of retrieved nodes, and then retrieves paths with the highest reliability scores. Answer Generation Stage: The retrieved paths are placed into prompts in ascending order of reliability scores and fed into an LLM for answer generation.
  • Performance: PathRAG consistently outperforms baselines (NaiveRAG, HyDE, GraphRAG, LightRAG) across various datasets and evaluation dimensions (comprehensiveness, diversity, logicality, relevance, coherence).

6. Code & Implementation (19:37 - 20:25)

  • PathRAG’s code is available on GitHub (BUPT-GAMMA/PathRAG).
  • The speaker highlights that PathRAG still requires a vector database storage (listing Neo4j, Oracle, Chroma, Milvus, TiDB, Mongo, AGE as supported options), which he sees as a slight drawback since it’s “falling back” to classical RAG components for initial node retrieval. However, it leverages open-source methodologies.

7. RAG in Medicine - A Strong Warning (20:25 - 25:49)

  • The speaker points to a recent paper from MIT, Stanford, and Duke (“Retrieval-augmented systems can be dangerous medical communicators”).
  • This paper argues that current RAG-based systems, when used in medical AI, can: Generate responses based on literal and narrow interpretations of queries. Reinforce patient presuppositions and biases. Decontextualize facts from source material. Produce misleading sentences. Generate results without an intuitive, pragmatic understanding of likely downstream consequences.
  • The speaker, as a theoretical physicist and not a medical doctor, emphasizes that he cannot personally evaluate the medical complexity or validity of these claims, but he highlights the strong warning from these reputable universities. He stresses that RAG in medicine still requires a lot of research and that he would personally prefer to consult a human doctor for medical conditions.

Conclusion (25:49 - 26:22) The video concludes by reiterating the fascinating development of RAG systems, with PathRAG being the current state-of-the-art in outperforming previous models. However, it also leaves the audience with a critical consideration regarding the ethical and practical implications of RAG, particularly in sensitive domains like medicine, where further research and caution are needed.

Graph View