SmartRAG - Intelligent Multimodal RAG System

A production-ready RAG system enabling intelligent conversations with documents, images, and audio files. Built with local-first AI models for complete privacy and offline operation.

Quick Start

# Standard deployment
docker-compose up -d

# Access at http://localhost:8501

Core Features

Multimodal Processing

• Documents: PDF, DOCX, TXT, MD with intelligent chunking
• Images: OCR + visual understanding via BLIP
• Audio: Automatic transcription with Whisper

Local AI Stack

• Ollama (Llama 3.1 8B) for generation
• Nomic Embed Text (768-dim) for embeddings
• ChromaDB for vector storage
• Complete offline operation

Production Ready

• Docker deployment with multi-stage builds
• Non-root user execution
• Health checks and auto-healing
• Resource management and monitoring
• Security hardening included

Technology Stack

Component	Technology
LLM	Llama 3.1 8B via Ollama
Embeddings	Nomic Embed Text (768-dim)
Vector DB	ChromaDB / FAISS
Vision	BLIP + CLIP + Tesseract OCR
Audio	OpenAI Whisper (base)
UI	Streamlit
Storage	SQLite3

Architecture

Installation

Docker (Recommended)

git clone https://github.com/itanishqshelar/SmartRAG.git
cd SmartRAG/docker

# Development
docker-compose up -d

# Production with full stack (PostgreSQL, Redis, Nginx)
docker-compose -f docker-compose.prod.yml up -d

Local Setup

# Install dependencies
pip install -r requirements.txt

# Install Ollama and models
ollama pull llama3.1:8b
ollama pull nomic-embed-text

# Install system dependencies
# macOS: brew install tesseract ffmpeg
# Ubuntu: apt-get install tesseract-ocr ffmpeg
# Windows: Download from GitHub releases

# Run application
streamlit run chatbot_app.py

Configuration

SmartRAG uses a single config.yaml with Pydantic validation:

models:
  llm_model: "llama3.1:8b"
  embedding_model: "nomic-embed-text"
  vision_model: "Salesforce/blip-image-captioning-base"
  whisper_model: "base"

vector_store:
  type: "chromadb"
  embedding_dimension: 768

processing:
  chunk_size: 1000
  chunk_overlap: 200
  ocr_enabled: true

generation:
  temperature: 0.7
  max_tokens: 2000
  context_window: 4096

Override via environment variables:

export SMARTRAG_LLM_MODEL=llama2:7b
export SMARTRAG_TEMPERATURE=0.5

Usage

Web Interface

1. Upload files via drag-and-drop
2. Ask questions about your content
3. View source documents inline
4. Manage chat history and files

Python API

from multimodal_rag.system import MultimodalRAGSystem

system = MultimodalRAGSystem()

# Ingest content
system.ingest_file("document.pdf")
system.ingest_file("screenshot.png")
system.ingest_file("recording.mp3")

# Query with context
response = system.query("Summarize the key points")
print(response.answer)

Batch Processing

# Process directories
results = system.ingest_directory("./docs/", recursive=True)
print(f"Processed {len(results)} files")

Project Structure

smartrag/
├── chatbot_app.py              # Streamlit application
├── config.yaml                 # Configuration
├── requirements.txt            # Dependencies
├── multimodal_rag/
│   ├── system.py              # RAG orchestrator
│   ├── processors/            # File type handlers
│   │   ├── document_processor.py
│   │   ├── image_processor.py
│   │   └── audio_processor.py
│   └── vector_stores/         # DB implementations
│       ├── chroma_store.py
│       └── faiss_store.py
├── docker/                    # Production deployment
│   ├── Dockerfile
│   ├── docker-compose.yml
│   └── docker-compose.prod.yml
└── tests/                     # Test suite

Deployment Options

Standard - All-in-one container with Ollama

docker-compose up -d

Lightweight - External Ollama on host

docker-compose -f docker-compose.lite.yml up -d

Production - Full stack with PostgreSQL, Redis, Nginx

docker-compose -f docker-compose.prod.yml up -d

Development

# Run tests
pytest tests/

# Code formatting
black multimodal_rag/ tests/

# Linting
flake8 multimodal_rag/ tests/

Performance

• Image size: 4.2GB
• Memory: 4-8GB recommended
• CPU: 2-4 cores recommended
• Startup time: ~90s (includes model downloads)
• Query latency: <3s typical

Security

• Local inference - no external API calls
• Non-root container execution
• File size limits enforced (50MB default)
• No privilege escalation
• Security headers in production setup

License

MIT License - see LICENSE file for details.

Acknowledgments

Built with ChromaDB, Ollama, Hugging Face Transformers, OpenAI Whisper, and Tesseract OCR.

---

SmartRAG - Local-first multimodal AI for document intelligence.