Add terminal-based multi-ecosystem support triage agent in code

[Abhinav9454]

Project Description for Multi-Domain Support Triage Agent

This terminal-based support triage agent automates the classification and response handling for tickets from three distinct ecosystems—HackerRank Support, Claude Help Center, and Visa Support—using only the provided offline support corpus.

Architecture and Approach
The agent follows a retrieval-augmented generation (RAG) pipeline:

Corpus Indexing: All HTML/Markdown documents from the data/ directory are chunked and embedded into a local vector database, enabling fast semantic retrieval without any network calls.

Ticket Processing: For each incoming ticket in support_tickets.csv, the agent retrieves the top-5 most relevant passages from the corresponding ecosystem's corpus.

Safe Triage via LLM: A carefully engineered system prompt enforces strict, document-grounded reasoning. The language model is instructed to:

Classify the request_type (FAQ, billing, fraud, etc.) and product_area.

Decide whether to reply or escalate based on a predetermined risk framework (fraud, missing documentation, internal data access, legal threats, etc.).

Produce a final response that contains only information explicitly present in the retrieved documents, or a mandatory escalation message.

Output Generation: The agent populates output.csv with exactly the fields required by the problem statement: request_type, product_area, escalation, retrieved_docs, and response.

Safety and Grounding

Zero Hallucination: The prompt forces the model to ground every statement in the retrieved documents; any unsupported query is immediately escalated.

High-Risk Escalation: Fraud, billing disputes, account compromises, security issues, and emotionally charged requests are always routed to a human.

Corpus-Only Knowledge: The agent never accesses the internet and relies exclusively on the provided data/ snapshots of HackerRank, Claude, and Visa support centres.

Key Features

Modular code that can be easily extended to new ecosystems or retrieval backends.

Transparent reasoning: the retrieved_docs field records the exact documentation URLs used for every reply, enabling full auditability.

Fully terminal-based: no GUI or web server; runs as a single Python script processing tickets sequentially.