HippoRAG

HippoRAG is an open-source Python framework from the OSU NLP Group at The Ohio State University. It is a retrieval and memory framework for large language models, inspired by the hippocampal indexing theory of human long-term memory. It indexes documents into a knowledge graph and supports retrieval and question answering over them. The current README covers HippoRAG 2, which builds on the original NeurIPS 2024 HippoRAG paper.

What It Is

HippoRAG is a retrieval-augmented generation (RAG) framework that orchestrates LLMs, knowledge graphs and the Personalized PageRank algorithm. The paper describes this as mimicking the different roles of the neocortex and hippocampus in human memory. The authors report that it outperforms existing RAG methods on multi-hop question answering by up to 20%. They also report that single-step retrieval is comparable to or better than iterative retrieval such as IRCoT while being 10-30 times cheaper and 6-13 times faster. The README says HippoRAG 2 improves associativity (multi-hop retrieval) and sense-making without sacrificing performance on simpler tasks. It also says HippoRAG 2 uses significantly fewer offline indexing resources than GraphRAG, RAPTOR and LightRAG.

Workflow and Model Support

Users install the package with pip, create a HippoRAG object with a save directory plus an LLM and an embedding model, index documents, and then run retrieval or rag_qa over queries. The framework supports the following backends:

• OpenAI and Azure OpenAI.
• Any OpenAI-compatible endpoint, such as a local vLLM server or a hosted gateway.
• Amazon Bedrock, including Bedrock Mantle.
• Local inference with vLLM or Transformers, and local embedding models such as NV-Embed-v2 and GritLM.

Embeddings are stored in local Parquet files by default. Qdrant, ChromaDB or Milvus can be selected instead. Indexes are tied to the models and configuration that built them, and an index with a mismatched identity is rejected rather than mixed.

Agent and MCP Integration

An optional hipporag-mcp server exposes a persistent index to MCP clients such as Claude Code, Claude Desktop, Cursor and VS Code. Its tools are retrieve, rag_qa, index, delete and index_stats. It runs over stdio or Streamable HTTP, and a read-only mode disables the index and delete tools. The repository also acts as a Claude Code plugin marketplace, with a bundled skill and an /hipporag:index-folder command.

Reproducing Experiments

A main.py entry point runs the paper experiments on sample and custom datasets, with an optional standard dense-retrieval baseline. The vLLM offline batch mode is documented as more than 3x faster for indexing than the online server. The HippoRAG 1 code is kept on a legacy branch.