A professional-grade Retrieval-Augmented Generation (RAG) application designed for auditing complex legal documents. Built with a robust containerized architecture, it moves beyond simple chat to provide Deep Contextual Audits of contracts using "Lost-in-the-Middle" mitigation strategies.

Sequence Diagram

• Brain: Gemini 1.5 Flash (via LangChain Google GenAI)
• Memory: Chroma Vector Cloud Database
• Interface: Streamlit (with Session State Memory)
• Infrastructure: Docker & Docker Compose

• Deep Contextual Audit: Implemented RecursiveCharacterTextSplitter with 2000-character chunks and high overlap to ensure legal definitions and multi-paragraph clauses remain semantically intact.
• Metadata Injection: Custom ingestion pipeline extracts document metadata (Filename, Date, Type) and injects it directly into vector embeddings to improve retrieval accuracy.
• Persistent Vector Memory: Integrated with Chroma to store and retrieve document embeddings efficiently, preventing re-indexing of duplicate files.
• Real-time Streaming: Implemented token-by-token response streaming using LangChain generators for a responsive user experience.
• Dockerized Architecture: Fully containerized microservices architecture ensuring consistent performance across development and production environments.

• Metadata Injection: Custom ingestion pipeline extracts document metadata (Filename, Date, Type) and injects it directly into vector embeddings.

Real-time ingestion tracking in the sidebar, confirming successful metadata injection.

Embeddings stored in Chroma DB

• Persistent Vector Memory: Integrated with Chroma to store and retrieve document embeddings efficiently.

• Real-time Streaming: Implemented token-by-token response streaming for a responsive experience.

• Docker and Docker Compose installed.
• A Google Gemini API Key.

Create a .env file in the root directory:

GOOGLE_API_KEY=your_gemini_api_key_here
CHROMA_KEYS=your_chroma_keys_here

Launch the entire stack from scratch to ensure a clean build:

# Build and start the containers
docker compose up --build -d

1. Access: Open the UI at http://localhost:8501.
2. Upload: Use the sidebar to upload PDF legal contracts.
3. Audit: Ask complex reasoning questions:
   • "What are the termination liabilities in Section 4.2?"
   • "Is there a non-compete clause that exceeds 12 months?"
4. Monitor: To check the logs in real-time:

   docker compose logs -f

legal-rag/
├── DATA/
│   └── UPLOADED/           # Persistent storage for PDF inputs
│       └── files.pdf
├── src/
│   ├── llm_model/          # Embedding & LLM initialization
│   │   ├── embedding.py
│   │   └── llm.py
│   ├── prompts/            # RAG chains & system templates
│   │   └── legal_template.py
│   ├── text_handler/       # Chunking & formatting logic
│   │   └── splitter.py
│   ├── utils/              # Helper functions (DB & Chat)
│   │   ├── chat_utils.py
│   │   └── db_utils.py
│   ├── database.py         # Vector store connection
│   ├── ingestion.py        # PDF extraction & metadata logic
│   └── retriever.py        # Two-stage retrieval pipeline
├── tests/                  # Evaluation & Testing suite
│   ├── deep_eval_model.py
│   ├── generate_eval_data.py
│   ├── rag_chain_eval.py
│   └── test_eval.py
├── app.py                  # Main Streamlit Entry Point
├── deepeval_test.py        # Automated testing script
├── Dockerfile              # Container definition
└── docker-compose.yml      # Orchestration config

To stop the application and clear the cache/volumes (useful for resetting the database):

docker compose down
docker system prune -a --volumes

1. Access: Open the UI at http://localhost:8501.

2. Upload: Use the sidebar to upload PDF legal contracts.

3. Audit: Ask complex reasoning questions:

   • "What are the termination liabilities in Section 4.2?"

   • "Is there a non-compete clause that exceeds 12 months?"

4. Monitor: To check the logs in real-time:

docker compose logs -f   

Found a bug or have a suggestion for the auditing pipeline? Please open an issue!