Multi-Domain Support Triage Agent

🏆 Built for the HackerRank Agentic AI Hackathon

A terminal-based support triage agent that handles support tickets across three ecosystems:

• HackerRank — https://support.hackerrank.com
• Claude — https://support.claude.com/en/
• Visa — https://www.visa.co.in/support.html

---

How It Works

For each support ticket the agent:

1. Identifies the domain — uses the company field or infers it via keyword scoring when company is None
2. Detects cross-domain confusion — if the ticket content belongs to a different domain than declared, it auto-corrects
3. Detects multiple requests — flags tickets containing more than one distinct issue
4. Runs safety pre-classification — catches prompt injection, malicious requests, sensitive triggers, and out-of-scope topics before the LLM
5. Retrieves relevant corpus — uses TF-IDF weighted scoring to extract the most relevant paragraphs from the scraped support corpus
6. Classifies and responds — sends the ticket + corpus snippet to an LLM with strict grounding rules
7. Validates output — sanitizes all fields before writing to CSV

---

Setup

1. Install dependencies

pip install -r requirements.txt

2. Set your API key

cp .env.example .env
# Edit .env and add your Groq API key (free at console.groq.com)

3. Scrape the support corpus (one-time)

python scrape_corpus.py

Creates corpus/hackerrank.txt, corpus/claude.txt, corpus/visa.txt from the live support sites.

4. Place your input CSV

Put support_tickets.csv in the project root.

5. Run the agent

python agent.py

Outputs output.csv and log.txt.

---

File Structure

.
├── agent.py                  # Main triage agent
├── scrape_corpus.py          # One-time corpus scraper
├── scrape_hackerrank_only.py # Re-scrape HackerRank + Claude only
├── requirements.txt          # Python dependencies
├── .env.example              # API key template
├── corpus/
│   ├── hackerrank.txt        # Scraped HackerRank support docs
│   ├── claude.txt            # Scraped Claude support docs
│   └── visa.txt              # Scraped Visa support docs
├── support_tickets.csv       # Input tickets
├── output.csv                # Agent predictions
└── log.txt                   # Full run transcript

---

Output Schema

Field	Allowed Values
status	replied, escalated
product_area	Support category string
response	User-facing answer grounded in corpus
justification	Internal reasoning + corpus source + retrieval confidence
request_type	product_issue, feature_request, bug, invalid

---

Key Features

TF-IDF Weighted Retrieval

Rare domain-specific terms (e.g. "chargeback", "proctoring", "impersonation") score higher than common words, ensuring the most relevant corpus paragraphs are selected for each ticket.

Multi-Layer Safety Pre-Classifier

Runs before the LLM and catches:

• Prompt injection — instruction overrides, identity manipulation, information extraction attempts, jailbreaks (categorized by type)
• Malicious requests — harmful code requests, system attacks
• Sensitive triggers — identity theft, fraud, security vulnerabilities, subscription changes
• Out-of-scope topics — irrelevant or off-domain queries
• Vague tickets — requests too ambiguous to classify

Cross-Domain Confusion Detection

Detects when a ticket's content belongs to a different domain than the declared company field and automatically corrects the routing.

Multi-Request Detection

Flags tickets containing multiple distinct requests, ensuring complex tickets are handled appropriately.

Corpus Source Citation

Every LLM-generated response includes the exact corpus URL and retrieval confidence score in the justification field, making all decisions fully traceable.

---

Escalation Logic

The agent escalates when:

• Fraud, identity theft, or account compromise is detected
• Account access loss or unauthorized access
• Security vulnerabilities or bug bounty reports
• Billing disputes with specific order IDs
• Legal threats or compliance concerns
• Prompt injection or adversarial manipulation detected
• Subscription cancellation or pausing requests
• Corpus does not contain sufficient information to answer safely

---

Hackathon

Built as a submission for the HackerRank Agentic AI Hackathon.