🧠 Multi-Domain Support Triage Agent

📌 Overview

This project implements a terminal-based AI support triage agent that processes customer support tickets across multiple ecosystems (e.g., HackerRank, Claude, etc.).

The agent:

• Classifies tickets into appropriate product areas
• Decides whether to respond or escalate
• Generates user-facing responses
• Provides a justification for its decision

---

⚙️ Setup Instructions

1. Extract the zip files

unzip code.zip
cd code

2. Create Virtual Environment

python -m venv venv
venv\Scripts\activate   # On Windows
# OR
source venv/bin/activate   # On Mac/Linux

3. Install Dependencies

pip install -r requirements.txt

4. Configure Environment Variables

Create a .env file in the root directory, and copy .env.example to .env and replace the values with your own API keys:

GEMINI_API_KEY=your_api_key
GROQ_API_KEY=your_api_key
OPENROUTER_API_KEY=your_api_key

EMBEDDING_MODEL=all-MiniLM-L6-v2
GOOGLE_MODEL=gemini-2.5-flash
GROQ_MODEL=llama-3.3-70b-versatile
OPENROUTER_MODEL=openai/gpt-4o-mini

5. Adjust the config.py file.

Make sure the path to the knowledge base is correct. You can change the paths as per your requirements. Also, you can change the models and keys as needed.

6. Run the Agent

python main.py

Output will be generated as:

OUTPUT_DIR/output.csv

---

🧩 Approach Overview

1. Ticket Processing Pipeline

Each ticket goes through the following stages:

a. Preprocessing

• Extract fields like company, issue type, and description
• Handle missing values (e.g., fallback for missing company)

b. Retrieval (RAG)

• Documents are filtered by company/domain
• Relevant knowledge is retrieved using vector search (FAISS / embeddings)

c. Product Area Classification

• A lightweight classifier maps tickets into categories such as:

  • Product Issue
  • Fraud
  • Inquiry
  • Bug

d. Safety & Escalation Logic

• Tickets are escalated if:

  • Sensitive keywords are detected (e.g., fraud, hacked, unauthorized access)
  • No relevant documents are retrieved
  • Confidence in response is low

e. Response Generation

• Uses LLM to generate:

  • Clear, user-facing response
  • Grounded strictly in retrieved documents

f. Justification Generation

• Short explanation describing:

  • Why the ticket was escalated OR
  • How the response was derived

---

📊 Output Format

The agent produces an output.csv with:

Column	Description
status	replied or escalated
product_area	Classified domain/category
response	Final user-facing reply
justification	Reasoning behind decision

---

🚀 Key Design Decisions

• RAG-based architecture for grounded responses
• Fail-safe escalation when knowledge is insufficient
• Modular pipeline (retrieval, classification, safety, generation)
• Deterministic structure for evaluation consistency