HippoRAG is a novel Retrieval-Augmented Generation (RAG) framework designed to imbue Large Language Models (LLMs) with human-like long-term memory capabilities. It enables continuous knowledge integration from external documents, enhancing associative reasoning and sense-making in complex contexts. The framework is optimized for cost and latency efficiency in online operations and requires fewer resources for offline indexing compared to graph-based RAG alternatives.

How It Works

HippoRAG integrates RAG with knowledge graphs and Personalized PageRank algorithms. It processes documents to extract factual triples, builds a knowledge graph, and then uses graph traversal techniques to improve retrieval. This approach allows LLMs to better recognize and utilize connections within new knowledge, mimicking human long-term memory functions for improved multi-hop retrieval and context integration.

Quick Start & Requirements

• Installation: conda create -n hipporag python=3.10, conda activate hipporag, pip install hipporag
• Environment Variables: CUDA_VISIBLE_DEVICES, HF_HOME, OPENAI_API_KEY (if using OpenAI models).
• Dependencies: Python 3.10, CUDA-enabled GPUs for local deployment.
• Quick Start: https://github.com/OSU-NLP-Group/HippoRAG#quick-start
• Reproducing Experiments: https://github.com/OSU-NLP-Group/HippoRAG#reproducing-our-experiments

Highlighted Details

• Achieves state-of-the-art performance in factual memory, sense-making, and associativity across multiple benchmarks.
• Supports OpenAI models and locally deployed vLLM instances for LLM inference.
• Offers efficient offline indexing compared to other graph-based RAG solutions.
• Includes comprehensive testing scripts for OpenAI and local deployments.

Maintenance & Community

• Developed by the OSU-NLP-Group, with contributions from Bernal Jiménez Gutiérrez, Yiheng Shu, and Yu Su.
• Contact: File an issue or contact the developers directly.
• Citation details provided for both HippoRAG and HippoRAG 2 papers.

Licensing & Compatibility

• The repository does not explicitly state a license in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

• Support for additional embedding models and embedding endpoints is listed as a TODO.
• Integration with vector databases is also a planned feature.
• Custom dataset formatting requires adherence to specific JSON structures for corpus and queries.