llama-4-multi-function-app

All-in-One Chat, OCR, RAG & Agentic AI App

A unified Streamlit interface that brings together conversational AI, document OCR, Retrieval-Augmented Generation (RAG), and multi-agent workflows - powered by CrewAI and the multimodal Llama-4 Scout model via Groq.

Features

Tab	What it does
Chat	Conversational AI with Llama 4 Scout multimodal model (text + image)
OCR	Extract text from images using Llama 4 Scout's vision capabilities
RAG	Upload PDFs, build a vector index, and ask questions grounded in your documents
Agentic AI	Multi-agent workflows powered by CrewAI (research, analysis, content generation)
RAG Evaluation	Evaluate RAG quality using RAGAS metrics (faithfulness, relevance, context precision)

API Keys (Both Free-Tier)

Key	Where to get it	Used for
GROQ_API_KEY	console.groq.com/keys	Primary LLM (Llama 4 Scout)
GOOGLE_API_KEY	aistudio.google.com/app/apikey	Fallback LLM for CrewAI + embeddings

Both APIs are free-tier. No credit card required.

Setup Instructions

Prerequisites

• Python 3.10 or higher
• pip (Python package installer)

Installation

1. Clone the repository:

   git clone https://github.com/genieincodebottle/generative-ai.git
   cd genai-usecases/llama-4-multi-function-app

2. Create a virtual environment:

   pip install uv    # if uv not installed
   uv venv
   .venv\Scripts\activate    # On Linux/Mac: source .venv/bin/activate

3. Install dependencies:

   uv pip install -r requirements.txt

   Note: This installs ~50 packages including CrewAI, RAGAS, sentence-transformers, and FAISS. First install may take 5-10 minutes.

4. Set up environment variables:

   cp .env.example .env

   Edit .env and add your keys:

   GROQ_API_KEY=your_groq_key_here     # Free-tier
   GOOGLE_API_KEY=your_google_key_here  # Free-tier

5. Run the app:

   streamlit run app.py

Usage Tips

• Start with Chat tab - simplest way to test your setup
• For RAG: Upload a PDF in the sidebar, wait for indexing, then ask questions
• For OCR: Upload an image and the model will extract text
• CrewAI agents take 2-5 minutes per run (this is normal - multiple agents collaborate)
• RAG Evaluation uses RAGAS metrics - requires both Groq and Google API keys

Tech Stack

• UI: Streamlit
• Primary LLM: Llama 4 Scout via Groq (free-tier)
• Fallback LLM: Google Gemini (free-tier, used in CrewAI)
• Vector Store: FAISS (in-memory, no database setup needed)
• Embeddings: HuggingFace sentence-transformers
• Multi-Agent: CrewAI framework
• Evaluation: RAGAS (Retrieval-Augmented Generation Assessment)

Troubleshooting

• "GROQ_API_KEY not set": Check your .env file exists in the project root
• CrewAI is slow: Multi-agent workflows involve multiple LLM calls - 2-5 minutes is normal
• RAGAS evaluation fails: Ensure both GROQ and GOOGLE API keys are set
• Memory issues: sentence-transformers downloads ~420MB model on first run