Lab 3: Chatbot vs ReAct Agent (Industry Edition)

Welcome to Phase 3 of the Agentic AI course! This lab focuses on moving from a simple LLM Chatbot to a sophisticated ReAct Agent with industry-standard monitoring.

🚀 Getting Started

1. Setup Environment

Copy the .env.example to .env and fill in your API keys:

cp .env.example .env

2. Install Dependencies

pip install -r requirements.txt

3. Directory Structure

• src/tools/: Extension point for your custom tools.

🏠 Running with Local Models (CPU)

If you don't want to use OpenAI or Gemini, you can run open-source models (like Phi-3) directly on your CPU using llama-cpp-python.

1. Download the Model

Download the Phi-3-mini-4k-instruct-q4.gguf (approx 2.2GB) from Hugging Face:

• Phi-3-mini-4k-instruct-GGUF
• Direct Download: phi-3-mini-4k-instruct-q4.gguf

2. Place Model in Project

Create a models/ folder in the root and move the downloaded .gguf file there.

3. Update .env

Change your DEFAULT_PROVIDER and set the path:

DEFAULT_PROVIDER=local
LOCAL_MODEL_PATH=./models/Phi-3-mini-4k-instruct-q4.gguf

🎯 Lab Objectives

1. Baseline Chatbot: Observe the limitations of a standard LLM when faced with multi-step reasoning.
2. ReAct Loop: Implement the Thought-Action-Observation cycle in src/agent/agent.py.
3. Provider Switching: Swap between OpenAI and Gemini seamlessly using the LLMProvider interface.
4. Failure Analysis: Use the structured logs in logs/ to identify why the agent fails (hallucinations, parsing errors).
5. Grading & Bonus: Follow the SCORING.md to maximize your points and explore bonus metrics.

🛠️ How to Use This Baseline

The code is designed as a Production Prototype. It includes:

• Telemetry: Every action is logged in JSON format for later analysis.
• Robust Provider Pattern: Easily extendable to any LLM API.
• Clean Skeletons: Focus on the logic that matters—the agent's reasoning process.

---

Sales Mail Agent Prototype

This repo now includes a local prototype for processing sales emails that request quotes or orders.

• Tools live in src/tools/sales_mail_tools.py.
• Local fixtures live in data/.
• The baseline chatbot lives in src/baseline/chatbot.py.
• The default demo provider is OpenAI via OPENAI_API_KEY.
• A deterministic mock provider lives in src/core/mock_provider.py, so the demo can still run without API keys.

Run the demo with OpenAI gpt-4o-mini:

python scripts/run_sales_mail_demo.py --provider openai --model gpt-4o-mini

Run one email only:

python scripts/run_sales_mail_demo.py --provider openai --model gpt-4o-mini --email-id email_001

Run offline with the mock provider:

python scripts/run_sales_mail_demo.py --provider mock

Run tests:

python -m pytest -q

The demo compares the chatbot baseline with the ReAct Agent and writes results to report/sales_mail_demo_results.json.

---

Happy Coding! Let's build agents that actually work.