ALEX: Adaptive Legal Explanation System

Existing legal Retrieval-Augmented Generation (RAG) systems rely on static workflows, resulting in resource misallocation for queries of varying complexity. To address this, we introduce ALEX, an adaptive retrieval architecture that dynamically matches reasoning depth to question complexity.

📔 Overview & Contributions

Figure 1: Overview of the ALEX.

The central innovation of ALEX lies in a two-stage process that first assesses the inherent complexity of a legal question and then routes it to a workflow specifically tailored to that level. By dynamically tailoring the reasoning process, ALEX not only improves accuracy on challenging legal benchmarks but also establishes an optimal balance between analytical depth and computational efficiency.

This paper makes the following contributions:

• A complexity-aware classification framework that automatically categorizes legal questions into three reasoning levels through dual-model validation and pseudo-labeling.
• An adaptive reasoning architecture featuring three specialized processing workflows that dynamically allocate computational resources based on question complexity.
• A confidence-driven pseudo-labeling algorithm tailored for legal domain data scarcity, effectively bootstrapping robust classifiers from minimal labeled data.
• Comprehensive experiments demonstrating substantial improvements over strong baselines while maintaining computational efficiency.

📦 Installation

Follow the steps below to download datasets, preprocess data, train models, and run inference.

Download Datasets

We use the following datasets. Please download them from Hugging Face and place them in the data/ folder:

• barexam_qa
• housing_qa

📥 Download Models

We use the following models. Please download them from Hugging Face and place them in the corresponding folders:

• Classification Model: flan-t5-large, place it in the flan-t5-large/ folder
• Embedding Model: bge-m3, place it in the bge-m3/ folder

Usage

Data Processing

Run the following preprocessing scripts to prepare the datasets for training and inference:

cd processed_data
python split.py
cd processed_data/classify
python label.py

Train Classification Model

Train the complexity classifier using the prepared labeled data:

cd processed_data/classify
python train.py

Extract PDF Text

Extract text content from the source PDF documents:

cd processed_data/passage
python processed_data/passage/pdf.py

Build Index

Extract text from source documents and build the vector database for retrieval:

pip install chromadb
cd processed_data/passage
python passage.py
python embedding2.py

Run Inference

Execute the main question-answering pipeline:

cd infer

python create_data.py

python deepseek_infer.py \
  --bar-exam-file data/bar_exam_labeled.jsonl \
  --housing-file data/housing_labeled.jsonl \
  --output-folder results