SaralPolicy - Local AI Insurance Analysis

Tagline: "Insurance ka fine print, ab saaf saaf."

Status: 🟢 POC/Demo (Remediation Complete) | Progress: 26/26 Issues Resolved (100%) | View Status | Implementation Report

🛠️ Tech Stack

🔧 OSS Frameworks

Purpose	License	Stars
RAG Evaluation (Faithfulness, Relevancy)	Apache 2.0	7k+
Metrics, Tracing, Observability	Apache 2.0	1.5k+
Background Task Queue (SQLite backend)	MIT	5k+
High-Quality Hindi Neural TTS (0.9B params)	Apache 2.0	AI4Bharat

🎯 What is SaralPolicy?

An AI-powered insurance document analysis system (POC/Demo) that uses local AI models to provide clear, easy-to-understand summaries in Hindi and English, with comprehensive guardrails and human-in-the-loop validation. All processing happens locally for complete privacy.

Status: 🟡 POC/Demo - See Status and Production Engineering Evaluation for details.

🖼️ Application Screenshots

Main Interface

Beautiful Material 3 design with drag & drop document upload

Real-time policy analysis with AI-powered insights

RAG Indexing & Knowledge Base

IRDAI knowledge base indexing with ChromaDB

Hybrid search combining BM25 + Vector embeddings

Policy Analysis Features

Comprehensive policy breakdown with key terms

Coverage details and exclusion highlights

Exclusion Detection

AI-powered exclusion identification with IRDAI citations

✨ Core Features

🚀 Performance & Processing

⚡ Optimized Document Parsing: Parallel PDF processing with ThreadPoolExecutor (4 workers)
📦 Smart Caching: Document and embedding caching with MD5 hashing
🔄 Batch Operations: Parallel embedding generation for faster RAG indexing
📊 Performance Monitoring: Real-time metrics tracking for all operations

🤖 AI & Intelligence

🤖 Local AI Analysis: Ollama + Gemma 2 2B model running locally for privacy
🧠 RAG-Enhanced: Retrieval-Augmented Generation with IRDAI knowledge base (39 chunks)
🔍 Hybrid Search: BM25 (keyword) + Vector (semantic) search with query caching
📤 Advanced Embeddings: nomic-embed-text (274MB) with connection pooling

🛡️ Safety & Quality

🔒 Advanced Guardrails: Input validation, PII protection, and safety checks
👥 Human-in-the-Loop: Expert review for low-confidence analyses
📊 RAGAS Evaluation: Faithfulness, relevancy, precision metrics (Apache 2.0)

📄 Document Processing

📁 Multi-Format Support: PDF, DOCX, and TXT files with drag & drop
🔑 Key Insights: Coverage details, exclusions, and important terms
❓ Interactive Q&A: Document-specific Q&A with IRDAI augmentation

🎨 User Experience

🎨 Modern UI: Beautiful Material 3 design with responsive layout
🌐 Bilingual: Results in both Hindi and English
📊 Text-to-Speech: High-quality Hindi neural TTS (Indic Parler-TTS)
🔒 100% Privacy: Complete local processing, no data leaves your machine

🚀 Quick Start

Prerequisites

OS: Windows, Linux, or macOS
Python: 3.9 or higher
RAM: Minimum 8GB (16GB recommended)
Disk Space: ~10GB (for models and virtual environment)

1. Setup Environment

# Clone the repository
git clone https://github.com/VIKAS9793/SaralPolicy.git
cd SaralPolicy

# Create virtual environment
python -m venv venv
venv\Scripts\activate  # Windows
# source venv/bin/activate  # Linux/Mac

# Upgrade pip
python -m pip install --upgrade pip

2. Install Dependencies

cd backend
pip install -r requirements.txt

3. Install and Setup Ollama

# Download and install Ollama from https://ollama.ai/download
# Pull required models
ollama pull gemma2:2b
ollama pull nomic-embed-text

# Start Ollama service (keep running in background)
ollama serve

4. Environment Configuration

# Copy environment template
copy .env.example .env  # Windows
# cp .env.example .env  # Linux/Mac

# Edit .env and add your HuggingFace token (optional, for Indic Parler-TTS)
# HF_TOKEN=hf_your_token_here

5. Index IRDAI Knowledge Base

python scripts/index_irdai_knowledge.py

6. Run the Application

python main.py

7. Open in Browser

Visit http://localhost:8000 to access the Material 3 web interface.

📁 Project Structure

SaralPolicy/
├── backend/
│   ├── main.py                          # Main FastAPI application
│   ├── requirements.txt                 # All dependencies (pinned versions)
│   ├── .env.example                     # Environment template (safe to commit)
│   ├── .env                             # Your secrets (NEVER commit)
│   ├── app/
│   │   ├── config.py                    # Validated configuration (Pydantic)
│   │   ├── dependencies.py              # Dependency Injection Container
│   │   ├── routes/                      # API Endpoints
│   │   ├── services/                    # Business Logic Services
│   │   │   ├── ollama_llm_service.py    # Local LLM via Ollama
│   │   │   ├── rag_service.py           # RAG with ChromaDB
│   │   │   ├── policy_service.py        # Policy analysis orchestration
│   │   │   ├── guardrails_service.py    # Input/output validation
│   │   │   ├── tts_service.py           # Text-to-speech orchestration
│   │   │   ├── indic_parler_engine.py   # Hindi neural TTS (Apache 2.0)
│   │   │   ├── rag_evaluation_service.py    # RAGAS evaluation
│   │   │   ├── observability_service.py     # OpenTelemetry metrics
│   │   │   └── task_queue_service.py        # Huey task queue
│   │   ├── middleware/                  # Security middleware
│   │   ├── prompts/                     # Versioned prompt registry
│   │   ├── models/                      # Data models
│   │   └── db/                          # Database layer
│   ├── tests/                           # All tests (197 passing)
│   ├── data/
│   │   ├── chroma/                      # ChromaDB persistent storage
│   │   └── irdai_knowledge/             # IRDAI regulatory documents
│   ├── scripts/                         # Utility scripts
│   ├── static/                          # Frontend assets (CSS, JS)
│   └── templates/                       # HTML templates
├── docs/
│   ├── SETUP.md                         # Installation guide
│   ├── TROUBLESHOOTING.md               # Common issues & solutions
│   ├── SYSTEM_ARCHITECTURE.md           # Detailed architecture
│   ├── STATUS.md                        # Current status
│   ├── setup/                           # Additional setup guides
│   ├── reports/                         # Technical reports
│   ├── adr/                             # Architectural decisions
│   └── product-research/                # Strategic documents (16 docs)
├── assets/                              # Screenshots and banners
├── .gitignore                           # Git ignore (includes .kiro/, .env)
├── README.md                            # This file
├── CONTRIBUTING.md                      # Contribution guidelines
└── LICENSE                              # MIT License

🔧 Technical Details

Core Infrastructure

Component	Technology	Version
Backend	FastAPI	0.115.6
AI Model	Ollama + Gemma 2 2B	Local
Embeddings	nomic-embed-text	274MB
Vector DB	ChromaDB	0.5.23
Validation	Pydantic	2.10.3
Database	SQLAlchemy + SQLite	2.0.36
UI	Material Design 3	HTML/CSS/JS

OSS Frameworks (No API Keys Required)

Framework	Purpose	License	Fallback
RAGAS	RAG evaluation (faithfulness, relevancy)	Apache 2.0	Heuristic scoring
OpenTelemetry	Metrics, tracing, observability	Apache 2.0	Built-in metrics
Huey	Background task queue (SQLite backend)	MIT	Synchronous execution
Indic Parler-TTS	High-quality Hindi neural TTS (0.9B params)	Apache 2.0	gTTS → pyttsx3

Text-to-Speech Performance

Hardware	~100 chars	~500 chars
CPU (default)	2-5 min	5-10 min
GPU (CUDA)	5-15 sec	15-45 sec
GPU (Apple M1/M2)	10-30 sec	30-90 sec

Note: Neural TTS on CPU is slow but produces high-quality Hindi speech. For faster responses, the system falls back to gTTS automatically.

🏗️ System Architecture

graph TB
    subgraph FE["🎨 FRONTEND"]
        UI["Material 3 Web UI"]
    end

    subgraph API["⚡ API LAYER"]
        FAST["FastAPI Server"]
    end

    subgraph CORE["🔧 CORE SERVICES"]
        DOC["Document Service"]
        RAG["RAG Service"]
        LLM["Ollama LLM<br/>gemma2:2b"]
        POLICY["Policy Service"]
    end

    subgraph STORAGE["💾 STORAGE"]
        CHROMA["ChromaDB"]
        IRDAI["IRDAI Knowledge<br/>39 chunks"]
    end

    subgraph SAFETY["🛡️ SAFETY"]
        GUARD["Guardrails"]
        EVAL["RAGAS Evaluation"]
        HITL["Human-in-the-Loop"]
    end

    subgraph AUX["📊 AUXILIARY"]
        TTS["TTS Service<br/>Indic Parler-TTS"]
        TRANS["Translation<br/>Argos Translate"]
    end

    UI --> FAST
    FAST --> DOC
    FAST --> POLICY
    POLICY --> RAG
    POLICY --> LLM
    RAG --> CHROMA
    IRDAI --> CHROMA
    POLICY --> GUARD
    POLICY --> EVAL
    EVAL --> HITL
    FAST --> TTS
    FAST --> TRANS

📖 Full Architecture: See SYSTEM_ARCHITECTURE.md

🧪 Testing

cd backend

# Run all tests (197 tests)
python -m pytest tests/ -v

# Run specific test categories
python -m pytest tests/test_security.py -v          # Security tests
python -m pytest tests/test_oss_frameworks.py -v    # OSS framework tests
python -m pytest tests/test_hallucination_detection.py -v  # Hallucination tests

# Run with coverage
python -m pytest tests/ --cov=app --cov-report=html

Test Categories

Category	Tests	Description
Configuration	23	Config validation, environment handling
Security	15	Input sanitization, PII protection, CORS
Error Paths	23	Graceful degradation, fallback handling
Hallucination	15	Detection accuracy, grounding validation
OSS Frameworks	28	RAGAS, OpenTelemetry, Huey integration
Integration	20+	End-to-end workflow tests
Total	197	All passing

🔒 Privacy & Security

100% Local Processing: All AI analysis happens on your machine
No Cloud APIs: Zero data sent to external services
No API Keys Required: Core functionality works offline
Advanced Guardrails: Input validation, PII protection, safety checks
HITL Validation: Expert review for quality assurance
Secure Secrets: .env file in .gitignore, never committed

Environment Variables

# backend/.env (NEVER commit this file)
HF_TOKEN=hf_your_token_here  # Optional: For Indic Parler-TTS
OLLAMA_HOST=localhost:11434  # Default Ollama endpoint

📚 Documentation

Quick Links

Changelog - Version history and release notes
Setup Guide - Installation and configuration
Troubleshooting - Common issues and solutions
System Architecture - Technical deep-dive
Status - Current project status
Ollama Setup - Local AI model setup

Product Research (16 Documents)

#	Document	Description
1	Executive Summary	Problem, solution, market opportunity (₹1,200 Cr)
2	Product Vision	Long-term vision and strategy
3	Business Case	Market sizing (515M policies), revenue models
4	Competitive Analysis	Competitor landscape, differentiation
5	Product Roadmap	Feature timeline, milestones
6	PRD	Functional requirements, user stories
7	Requirements	Technical specifications
8	Architecture	System design, data flow
9	Explainability	AI transparency, evaluation metrics
10	User Journey	UX flows, interaction design
11	Privacy & Compliance	IRDAI, DPDP Act 2023
12	Testing Strategy	QA plans, coverage
13	Go-to-Market	Launch plan, pricing
14	Risk Register	Risk identification, mitigation
15	Ethical AI	AI ethics, bias mitigation
16	Metrics & KPIs	Success metrics, North Star

📊 API Endpoints

Endpoint	Method	Purpose
`/`	GET	Serve frontend UI
`/upload`	POST	Upload policy document
`/analyze`	POST	Analyze uploaded policy
`/rag/ask`	POST	Ask question via RAG
`/rag/stats`	GET	RAG service statistics
`/tts`	POST	Generate audio summary
`/translate`	POST	Hindi ↔ English translation

🎯 POC Goals

Validate Core Concept: Does local AI analysis provide value?
Test Accuracy: Are summaries and insights accurate?
User Experience: Is the Material 3 interface intuitive?
Guardrails Effectiveness: Are safety checks working?
HITL Integration: How often is expert review needed?
Market Fit: Is there demand for privacy-first insurance analysis?

👤 Author

Vikas Sahani
Product Manager & Main Product Lead

📧 Email: vikassahani17@gmail.com
💼 LinkedIn: linkedin.com/in/vikas-sahani-727420358
🐙 GitHub: @VIKAS9793

Engineering Team:

Kiro (AI Co-Engineering Assistant)
Antigravity (AI Co-Assistant & Engineering Support)

🤝 Contributing

Contributions are welcome! Please read CONTRIBUTING.md before submitting.

# Quick start for contributors
git checkout -b feature/AmazingFeature
git commit -m 'feat: add AmazingFeature'
git push origin feature/AmazingFeature
# Open a Pull Request

📄 License

This project is licensed under the MIT License - see LICENSE for details.

🙏 Acknowledgments

IRDAI for insurance regulatory guidelines
Ollama for local LLM infrastructure
Google for Gemma models
ChromaDB for vector database
FastAPI community for excellent framework
AI4Bharat for Indic Parler-TTS Hindi speech synthesis
HuggingFace for model hosting and transformers library

📚 Citations

If you use SaralPolicy or its components in your research, please cite:

Indic Parler-TTS (Hindi Text-to-Speech)

@inproceedings{sankar25_interspeech,
  title     = {{Rasmalai : Resources for Adaptive Speech Modeling in IndiAn Languages with Accents and Intonations}},
  author    = {Ashwin Sankar and Yoach Lacombe and Sherry Thomas and Praveen {Srinivasa Varadhan} and Sanchit Gandhi and Mitesh M. Khapra},
  year      = {2025},
  booktitle = {{Interspeech 2025}},
  pages     = {4128--4132},
  doi       = {10.21437/Interspeech.2025-2758},
}

@misc{lacombe-etal-2024-parler-tts,
  author = {Yoach Lacombe and Vaibhav Srivastav and Sanchit Gandhi},
  title = {Parler-TTS},
  year = {2024},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/huggingface/parler-tts}}
}

@misc{lyth2024natural,
  title={Natural language guidance of high-fidelity text-to-speech with synthetic annotations},
  author={Dan Lyth and Simon King},
  year={2024},
  eprint={2402.01912},
  archivePrefix={arXiv},
}

Model: ai4bharat/indic-parler-tts (Apache 2.0, 0.9B parameters)

📞 Support

📧 Email: vikassahani17@gmail.com
🐛 Issues: GitHub Issues
💼 LinkedIn: Vikas Sahani

Found a security issue? Please email vikassahani17@gmail.com instead of using the public issue tracker.

Made with ❤️ to make insurance understandable for everyone
"Because everyone deserves to understand what they're paying for."

This is a POC/demo system. For production deployment, see the Production Engineering Evaluation and Remediation Plan.

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
assets		assets
backend		backend
docs		docs
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md

License

VIKAS9793/SaralPolicy

Folders and files

Latest commit

History

Repository files navigation