LegalGraphRAG
An evaluation framework and reference implementation for a multi-agent, graph-based retrieval-augmented generation method aimed at legal judgment prediction.
LegalGraphRAG is the official code for a paper on multi-agent graph retrieval-augmented generation for legal reasoning, accepted at ACL 2026. It provides a reproducible pipeline for the paper's main experiment, computing charge, law-article and prison-term prediction metrics against Chinese criminal law datasets, using a graph built from more than fourteen thousand cases.
It supports several underlying models, including Qwen, DeepSeek, GPT, InternLM, GLM and Gemma, and its main experiment run requires a CUDA-capable environment with enough memory to load a Qwen3-8B model per worker, along with a locally running embedding service. The repository includes the full case corpus, its construction commands, and the evaluation protocol needed to reproduce the reported results.