🧿 Vision Agent (Local Image Analysis)

A lightweight local AI vision agent built with Streamlit + LangGraph + Ollama that analyzes images and answers user questions using a multi-step pipeline.

Features

• Upload any image (JPG, PNG, WebP, GIF)

• Multi-step reasoning pipeline:

  • Vision → image description (LLaVA)
  • Research → reasoning over description
  • Writer → clean final answer

• Fully local (no API costs)

• Clean modern UI with Streamlit

---

Architecture

User Input
   │
   ▼
[ Vision Model (llava-phi3) ]
   │
   ▼
[ Research Agent (llama3.2) ]
   │
   ▼
[ Writer Agent (llama3.2) ]
   │
   ▼
Final Answer

Built using LangGraph state machine.

---

Requirements

• Python 3.9+
• Ollama installed → https://ollama.com
• 8GB RAM recommended

---

Setup

# Install dependencies
pip install -r requirements.txt

# Start Ollama
ollama serve

# Pull models
ollama pull llava-phi3
ollama pull llama3.2:1b

---

Run App

streamlit run frontend.py

---

Project Structure

.
├── frontend.py   # Streamlit frontend
├── backend.py    # LangGraph pipeline
└── README.md

---

How It Works

1. Image is converted to base64
2. Sent to LLaVA (vision model) via Ollama
3. Output is analyzed by LLM (llama3.2)
4. Final answer is generated and displayed

---