LegalRAG: Agentic AI for Explainable Legal Strategy ⚖️

Live Demo: https://explainable-legal-rag.streamlit.app/

📜 Overview

LegalRAG is an advanced Agentic AI system designed to act as an intelligent co-pilot for legal defense strategy. Unlike standard chatbots that simply "guess" answers or basic RAG systems that blindly retrieve text, LegalRAG uses a Reason+Act (ReAct) cognitive architecture.

It autonomously analyzes case files, identifies the correct jurisdiction (Civil vs. Military), selects the appropriate legal statutes using discrete tools, and generates a procedurally sound, step-by-step legal roadmap.

---

🚀 Key Features

1. 🧠 Agentic Orchestration (LangGraph)

Powered by LangGraph, the system doesn't follows a linear script. It thinks before it acts.

• Dynamic Routing: Automatically detects if a case involves Military Personnel and routes queries to the Army Act instead of Civil Code.
• Chain of Thought: You can see the agent's reasoning process (e.g., "The user is asking about a court-martial. I should check Section 63 of the Army Act.").

2. 📄 High-Fidelity Ingestion (Unstructured.io)

Legal documents are complex—filled with multi-column layouts, tables, and marginalia.

• Layout Awareness: We use UnstructuredPDFLoader to parse PDFs, ensuring that tabular data (like schedules of fines or dates) is preserved as structured information, not jumbled text.
• Semantic Chunking: Text is split in a way that preserves the meaning of legal clauses.

3. 🛠️ specialized Tool Use

The agent has access to a secure "Toolkit" to prevent hallucination:

• retrieve_from_case_file: Reads the specific facts of the uploaded PDF case.
• retrieve_from_cpc: Searches the Code of Civil Procedure (CPC) 1908.
• retrieve_from_army_code: Searches the Army Act, 1950 and Rules.

4. 🔍 Explainability & Traceability

Trust is critical in law.

• Trace Logs: The sidebar displays the exact active "Thought Process" and JSON outputs of every tool call.
• Citations: Every claim is backed by a specific Section or Page Number from the source document.

---

🆚 Comparison: Why Agentic AI?

Feature	Standard LLM (ChatGPT/GPT-4)	Standard RAG	LegalRAG (Agentic)
Data Source	Training Data (Often Outdated)	Static Document Search	Dynamic Tool Selection
Reasoning	Implicit / Hazy	None (Keyword Matching)	Explicit Multi-Step (ReAct)
Hallucination	High Risk (Invents Laws)	Medium (Wrong Context)	Near Zero (Grounded)
Input Parsing	Text Paste (Loses Formatting)	Basic PDF Readers	Unstructured.io (Layout Aware)
Transparency	Black Box	"Sources" Link	Full Execution Trace

---

📊 Evaluation Results

We rigorously tested the pipeline against complex, multi-jurisdictional synthetic scenarios.

• 100% Keyword Recall: The agent successfully identified all key legal concepts (e.g., "Section 63", "Tribunal", "Appeal Procedure") in every test case.
• 100% Routing Accuracy: It correctly distinguished when to apply Military Law vs. Civil Law in 3/3 complex test scenarios.
• Latency: ~100s average response time. (We prioritize deep reasoning correctness over speed).

---

🛠️ Installation & Local Setup

1. Clone the Repository

   git clone https://github.com/yourusername/LegalRAG.git
   cd LegalRAG

2. Install Dependencies

   pip install -r requirements.txt

3. Set API Keys Create a .env file:

   GROQ_API_KEY=your_groq_api_key_here
   HUGGINGFACEHUB_API_TOKEN=your_hf_token_here

4. Ingest Data (Build Vector Store)

   python LegalRAG/data_ingestion.py

5. Run the App

   streamlit run LegalRAG/app.py

---

🧪 Running Investigations

To run the automated evaluation suite:

python LegalRAG/evaluate_pipeline.py

This will generate a pipeline_evaluation_report.md with detailed performance metrics.

---

👨‍💻 Tech Stack

• Orchestration: LangChain, LangGraph
• LLM: Llama 3 / GPT-OSS (via Groq)
• Vector Store: FAISS
• Embeddings: HuggingFace (all-MiniLM-L6-v2)
• ETL/Ingestion: Unstructured.io
• Frontend: Streamlit

Contributors

• Parthiv Godrihal
• Nilay Jain
• Manan Bansal