👗 StyleAI — Agentic AI Fashion Stylist

EAG V3 · Assignment 3 — Agentic AI with Multi-Step Tool Calling

---

🎬 Demo Video

---

🧠 What is StyleAI?

StyleAI is a fully Agentic AI personal fashion stylist built with Google Gemini and Streamlit. It doesn't just answer a question — it reasons, calls external tools, synthesises real-world data, and delivers a personalised outfit recommendation.

Given your event, body, and style preferences, the agent:

1. Checks live weather at your event location
2. Fetches latest fashion news & trend headlines
3. Retrieves seasonal colour palette for your skin tone
4. Looks up dress code rules for your event type
5. Pulls outfit inspiration images (Pexels) for a visual moodboard
6. Synthesises everything into a final styled recommendation

---

🤖 Agentic Loop — How It Works

This is the core of the assignment. StyleAI calls Gemini multiple times in a loop, where every LLM response can trigger a tool call, whose result is fed back into the next prompt — until the agent reaches a final answer.

User Query
    │
    ▼
┌─────────────────────────────────────────────┐
│  LLM Call (Gemini)                          │
│  → Reasons about what data it needs         │
│  → Outputs: TOOL_CALL: get_weather | ...    │
└────────────────┬────────────────────────────┘
                 │ Tool detected
                 ▼
        🛠️ Tool Executed
        (real API call)
                 │
                 ▼
┌─────────────────────────────────────────────┐
│  Tool Result appended to conversation       │
│  → New prompt: "Tool returned X. Continue" │
└────────────────┬────────────────────────────┘
                 │
                 ▼
        🔁 Repeat (up to 8 steps)
                 │
                 ▼
        ✅ FINAL RECOMMENDATION

Every single step is displayed live in the UI — users see the agent's thinking, each tool call with its arguments, and the raw tool result, before the final answer appears.

---

🛠️ 5 Custom Tool Functions

#	Tool	API	What it does
1	get_weather	OpenWeatherMap	Fetches live temperature, humidity, wind, description + fabric advice for the event city & date
2	get_fashion_news	NewsAPI	Retrieves the latest fashion & trend headlines matching a query
3	get_outfit_images	Pexels	Searches for outfit inspiration photos; feeds the visual moodboard
4	get_color_palette	Internal	Derives a seasonal colour palette (Spring/Summer/Autumn/Winter) + skin-tone tips
5	get_dress_code_guide	Internal	Returns detailed dress-code rules, do's & don'ts for 8 event types

---

🌐 External APIs Used

1. Google Gemini (gemini-2.5-flash)

• Provider: Google AI Studio
• Used for: The core LLM — called multiple times in the agentic loop to reason, decide which tool to call next, and produce the final outfit recommendation
• Free tier: 15 RPM on gemini-2.0-flash, 5 RPM on gemini-2.5-flash
• Docs: aistudio.google.com

2. OpenWeatherMap

• Used for: get_weather tool — live temperature, humidity, wind speed, and weather description for any city
• Endpoint: GET /weather (current weather, free tier)
• Free tier: 60 calls/minute, no credit card required
• Docs: openweathermap.org/api

3. NewsAPI

• Used for: get_fashion_news tool — fetches latest news articles matching a fashion/trend keyword query
• Endpoint: GET /v2/everything
• Free tier: 100 requests/day (developer plan)
• Docs: newsapi.org

4. Pexels

• Used for: get_outfit_images tool — searches for high-quality outfit photos to populate the Pinterest-style moodboard
• Endpoint: GET /v1/search with Authorization: <key> header
• Free tier: 200 requests/hour, no credit card required
• Docs: pexels.com/api

---

📋 LLM Conversation Log Format

Each agent run is saved as a timestamped JSON in logs/. Here's the structure:

{
  "timestamp": "2026-04-24T18:30:00",
  "form_data": { "event_type": "Wedding", "event_city": "Mumbai", ... },
  "agent_steps": [
    { "step": 1, "type": "thinking",    "content_preview": "I need weather data..." },
    { "step": 1, "type": "tool_call",   "tool_name": "get_weather", "tool_args": {...} },
    { "step": 1, "type": "tool_result", "tool_name": "get_weather", "content_preview": "{...}" },
    { "step": 2, "type": "thinking",    "content_preview": "Weather is 32°C. Now fashion news..." },
    ...
    { "step": 5, "type": "final",       "content_preview": "### FINAL OUTFIT RECOMMENDATION..." }
  ],
  "final_answer": "### FINAL OUTFIT RECOMMENDATION\n..."
}

📄 Full session logs (including all LLM responses) are viewable in the 📋 Session Logs tab.

---

📁 Project Structure

• app.py — Streamlit UI entry point; defines all 4 tabs, custom CSS theming, and the sidebar

• requirements.txt — Python package dependencies

• .env — API keys (not committed to version control)

• code/ — Core application package

  • agent.py — Agentic loop: calls Gemini multiple times, detects tool calls, feeds results back, retries on rate limits
  • tools.py — 5 custom tool functions plus the Pexels moodboard image helper
  • config.py — Centralised API key loader via python-dotenv
  • logger.py — Saves and reads JSON session logs

• prompts/ — All LLM prompt templates

  • stylist_prompt.py — System prompt, user query template, and tool result template

• logs/ — Auto-saved JSON session logs, one file per agent run (e.g. session_20260424_183000.json)

---

🖥️ UI — 4 Tabs

Tab	Purpose
👗 Inputs	Three sections: Event Details · Personal Details · Style Preferences
✨ Outfit Recommendation	Live agentic reasoning chain → final recommendation → auto-generated moodboard
📌 Style Inspiration	Pinterest-style masonry board · Pexels search · Seasonal colour palette
📋 Session Logs	Browse, inspect and replay all past agent runs as JSON

---

⚙️ API Keys Required

Key	Required	Free Tier	Get It
GEMINI_API_KEY	✅	Yes	aistudio.google.com
OPENWEATHER_API_KEY	✅	Yes	openweathermap.org/api
NEWS_API_KEY	✅	Yes	newsapi.org/register
PEXELS_API_KEY	✅	Yes	pexels.com/api

---

🚀 Setup & Run

# 1. Clone / navigate to the project
cd "Assignment-3-StyleAI"

# 2. Install dependencies
pip install -r requirements.txt

# 3. Set up environment
copy .env.example .env
# → Open .env and fill in your API keys

# 4. Run
streamlit run app.py

---

🔑 Assignment Checklist

Requirement	Status
LLM called multiple times in a loop	✅ Up to 8 steps
Full conversation history passed each call	✅ history[] accumulates all turns
At least 3 custom tool functions	✅ 5 tools
Agent reasoning chain displayed in UI	✅ Thinking / Tool Call / Tool Result per step
External API usage	✅ OpenWeatherMap · NewsAPI · Pexels
Final answer clearly shown	✅ ### FINAL OUTFIT RECOMMENDATION
Session logs saved	✅ JSON in logs/
Rate-limit retry logic	✅ Auto-retry with parsed wait time

---

📝 Sample LLM Log (copy-paste for submission)

After running the agent, go to 📋 Session Logs tab → select the session → expand Raw JSON Log → copy and paste the full JSON for your submission.

---

Built with ❤️ using Google Gemini · Streamlit · OpenWeatherMap · NewsAPI · Pexels