Table of Content
AI Information
AI Tutorial
AI Agent
Latest Information
Hot Information
LLMs Information
Technical Frontiers
Industry Applications
Prompt Tutorial
AI Programming
AI Video Generation
AI Audio Generation
Knowledge Base Setup
AI Development Guides
Industry Tutorials
Practical Cases
AI Resource Library
Agent Information
Agent Setup
Agent Training

HippoRAG 2 released, GraphRAG abdicates~

Written by
Silas Grey
Updated on:July-14th-2025
Recommendation

HippoRAG 2 leads to a new breakthrough in RAG systems, taking simulation of human long-term memory a step further.

Core content:
1. The innovation of the HippoRAG 2 framework and its improvement over existing RAG systems
2. Evaluation and performance of HippoRAG 2 in key dimensions
3. Comparison of HippoRAG 2 with baseline methods and details of performance improvement

Yang Fangxian
Founder of 53AI/Most Valuable Expert of Tencent Cloud (TVP)
To address the limitations of existing retrieval-augmented generation (RAG) systems in simulating the dynamics and associativity of human long-term memory , a new framework HippoRAG 2 is proposed and will be open sourced.
Continuous learning capabilities are evaluated on three key dimensions: fact memory, perceptual construction, and relevance. HippoRAG 2 outperforms other methods ( RAPTOR, GraphRAG, LightRAG, HippoRAG ) in all benchmark categories, bringing it closer to true long-term memory systems.
The core idea of ​​the HippoRAG 2 framework : HippoRAG 2 is based on HippoRAG's personalized PageRank algorithm, which pushes the RAG system closer to the effect of human long-term memory through deep paragraph integration and more effective use of online LLM.

Offline indexing:

  • We use LLM to extract triples from paragraphs and integrate them into an open knowledge graph (KG).
  • Detect synonyms through embedding model and add synonym edges in KG.
  • Combine the original paragraph with the KG to form an open KG containing concepts and contextual information.
Online search:
  • Use embedding models to link queries with triples and paragraphs in the KG and determine seed nodes for graph search.
  • The retrieved triples are filtered through LLM to keep the relevant triples.
  • A personalized PageRank algorithm is applied for context-aware retrieval, ultimately providing the most relevant paragraphs for downstream question-answering tasks.
      Baseline methods: including classic retrievers (BM25, Contriever, GTR), large embedding models (GTE-Qwen2-7B-Instruct, GritLM-7B, NV-Embed-v2), and structure-enhanced RAG methods ( RAPTOR, GraphRAG, LightRAG, HippoRAG ).
      Evaluation metrics: The question answering task uses the F1 score, and the retrieval task uses passage recall@5.
      Performance improvement: HippoRAG 2 outperforms other methods on all benchmark categories, with an average F1 score of 7 percentage points higher than standard RAG, especially on the associative memory task.
      A HippoRAG 2 pipeline example
      https://github.com/OSU-NLP-Group/HippoRAG From RAG to Memory: Non-Parametric Continual Learning for Large Language Modelshttps://arxiv.org/pdf/2502.14802

      Related PostsView More
      HR Tips: SOP for Outstanding Employees
      Explore the efficient method of SOP combing for high-performing employees! With the help of a large model knowledge base, use prompt word techniques and frameworks to quickly extract key actions from interview recordings. Detailed analysis of each module requirement, such as clarifying work goals and background, breaking down work processes, etc. Let you easily master the process and improve...
      Iris Vance
      Context compression: the core step of RAG information extraction
      In-depth interpretation of RAG technology, focusing on the core step of context compression. Traditional RAG methods have problems such as noise interference and context window limitation, so context compression technology came into being. It can streamline documents, remove noise, increase information density, and improve answer quality and efficiency. To learn more about RAG technology and...
      Audrey Miles
      Problems encountered when building code execution tools based on Function Calling
      In-depth analysis of the difficulties in building code execution tools based on Function Calling, and exploration of innovative practices in intelligent agent development platforms. The article details how the model has the ability to call tools and how the backend handles complex instructions. It also shares model fine-tuning techniques and methods, such as online search and execution of...
      Silas Grey