xmudeeplit/legalgraphrag

Purpose and entry points

LegalGraphRAG is a Python evaluation framework for criminal legal judgment prediction. It builds a feature/case graph, retrieves graph-grounded legal context, and uses LLM agents to predict charges, applicable law articles, and imprisonment. The operational entry point is run.py; offline scoring is evaluation/evaluate_results.py.

Orchestration and execution

run.py loads dotenv configuration, selects a model/runtime, optionally builds or loads a serialized graph, distributes cases across requested GPU devices, and writes combined per-dataset predictions and run statistics. The central application coordinator is core/LegalGraphRAG.py, which connects preprocessing, retrieval, prompts, and judgment stages.

Grounding: run.py, core/LegalGraphRAG.py, configs/main.env.

Case preprocessing and feature extraction

Raw or evaluation cases are segmented, pre-judged, and converted into structured legal features used as graph nodes/retrieval keys. The preparation pipeline can generate sampled CAIL inputs, case features, law-dependency hints, and merged law-to-crime resources before inference.

Grounding: core/preprocess/case_seg.py, core/preprocess/get_features.py, core/preprocess/preJudge.py; orchestration scripts: scripts/prepare_data.py, scripts/prepare_case_features.py.

Graph construction and persistence

The graph layer converts feature-enriched historical cases and law resources into a retrievable graph database. It uses embeddings during construction/retrieval and persists the graph as a pickle artifact so parallel inference workers can reload a common read-only-ish graph rather than rebuild it. Rebuilds are required when graph-construction model/configuration changes.

Grounding: core/graph_construct/feature_graph.py, core/graph_construct/graph_db.py, datas/main_experiment/cases_with_feature_big.json.

Retrieval-to-judgment pipeline

For each target case, the system uses structured features to retrieve relevant precedent/law graph context, then routes that evidence through specialized legal decision agents. Crime and law judgments are separate boundaries, allowing charge reasoning and statutory-article reasoning to be independently produced before the final result schema is emitted.

Grounding: core/LegalGraphRAG.py, core/judge/judge_crime.py, core/judge/judge_law.py.

Model and prompt boundary

A common model abstraction supports local Hugging Face Transformers models and OpenAI-compatible API models. Model-specific implementations encapsulate loading/inference details; prompt packages separate graph, retrieval, preprocessing, and judge prompt families, with Chinese/English selection configured externally.

Grounding: core/models/base.py, core/models/transformers_base.py, core/models/openai_base.py; examples: core/models/transformers/qwen3_model.py, core/models/openai/gpt4o_mini.py.

Data assets, infrastructure, and evaluation

The main experiment consumes a 14,049-case graph corpus plus CAIL and CMDL evaluation inputs, criminal-law records, and crime-category metadata. It requires Python 3.10+, CUDA/PyTorch for local large models and multi-GPU execution; embeddings come from an externally running Ollama-compatible endpoint (default localhost:11434/api/embed, bge-m3). API-backed models additionally require configured credentials. Metrics evaluate charge/law exact match and Micro-F1, plus sentence exact match and month MAE.

Grounding: datas/main_experiment/crime_data_CAIL_small.json, datas/main_experiment/crime_data_CMDL_small.json, datas/main_experiment/criminal_law_processed.json, evaluation/evaluate_results.py.