CatRAG
A retrieval-augmented generation framework built on the HippoRAG 2 architecture that turns a static knowledge graph into a query-adaptive navigation structure, reproduced with the prompts and dataset from its ACL 2026 paper.
CatRAG is a retrieval-augmented generation framework, built on the HippoRAG 2 architecture, that adapts graph traversal to each query instead of using a fixed static graph. It applies symbolic anchoring to inject a weak seed concept related to the query, query-aware dynamic edge weighting to amplify relevant paths and prune irrelevant ones, and key-fact passage weight enhancement to boost edges connected to relevant context, steering a random walk toward the evidence chain a multi-hop question actually needs.
This repository provides a reproduced implementation of CatRAG together with the prompts and the HoVer dataset used in the paper's experiments. It supports OpenAI and OpenAI-compatible chat and embedding endpoints by default, with optional extras for local Transformers or GritLM embeddings, Amazon Bedrock providers, and offline information extraction through vLLM, the last of which is intended primarily for Linux GPU environments.
It is written in Python, requires Python 3.10 or newer, and its paper was accepted to the Findings of ACL 2026.