This repository implements a high-precision legal AI system designed for the Indian legal domain. It utilizes a hybrid architecture that combines an Agentic RAG pipeline for complex reasoning with QAT-optimized fine-tuning for domain-specific comprehension.
- Agentic RAG Pipeline: Built using LangGraph, the system executes multi-step query decomposition and conditional refinement loops. Instead of a single retrieval step, the agent decomposes legal queries, executes targeted searches, evaluates sufficiency, and iterates until the evidence context is sufficient to answer the query accurately.
- Domain-Specific Optimization: The backbone includes a Gemma4-E2B QAT-fine-tuned model. We employed Parameter-Efficient Fine-Tuning (PEFT) and QLoRA to align model performance with complex Indian legal datasets while maintaining computational efficiency.
- Evaluation Metrics: The pipeline achieves a 0.889 BERTScore-F1, consistently outperforming zero-shot baselines by effectively anchoring responses in retrieved legal statutes and precedents.
agentic_rag_core.py: The heart of the system. Implements the LangGraph state machine, nodes (decomposer, planner, retriever, evaluator, generator), and conditional routing logic.Agentic_rag_app.py: A Streamlit-based frontend for interacting with the agent, complete with reasoning trace visualization.Qwen3_4B_Finetune.ipynb: Notebook documenting the fine-tuning workflow (PEFT/QLoRA) applied to enhance domain vocabulary and reasoning.evaluate_rag.py: Scripts used for performance benchmarking and BERTScore calculation.
The system addresses the common pitfalls of vanilla RAG in the legal domain (e.g., retrieval failure on multi-faceted legal questions) by forcing the model to verify evidence sufficiency before final synthesis. This iterative feedback loop significantly reduces hallucinations and increases citation accuracy.
Developed as part of an effort to modernize access to Indian legal information through high-performance natural language understanding and agentic workflows.