ZenithSupport AI

A multi-agent RAG system designed to resolve customer support tickets with grounding, citations, and safety controls.

Live Demo (Wait for dependencies to install): zenithsupport-ai.streamlit.app Video Walkthrough: Watch the Technical Overview

Overview

This system uses a 4-agent orchestration pattern to process incoming support tickets. It retrieves relevant policies from a FAISS vector store, drafts resolutions based strictly on those policies, and validates the output for compliance and citations.

Core Agents

1. Triage Agent: Classifies the issue and identifies missing metadata.
2. Policy Retriever Agent: Fetches relevant semantic chunks from the policy corpus.
3. Resolution Writer Agent: Drafts a customer-ready response grounded in the retrieved text.
4. Compliance Agent: Verifies citations and ensures no unsupported claims exist.

Tech Stack

• Framework: LangChain (Multi-agent orchestration)
• Vector Store: FAISS (Local)
• Embeddings: HuggingFace all-MiniLM-L6-v2 (Local)
• LLM: Groq Llama 3.3 70B (High-speed reasoning)
• Data: Markdown-based "Heavy-Duty" policy corpus (27,000+ words, 14 documents).

Project Structure

• ecommerce_agent/data/policies/: Synthetic policy documents.
• ecommerce_agent/src/: Core logic (Ingestion, Engine).
• ecommerce_agent/tests/: Evaluation set of 20+ scenarios.
• ecommerce_agent/evaluate.py: Main evaluation script.

Input Format

The system accepts two primary inputs:

1. Ticket Text (Free-form)

A string containing the customer's query or issue.

2. Order Context (Structured JSON)

A dictionary containing the following mandatory fields:

• order_date: (String) Date the order was placed (YYYY-MM-DD).
• delivery_date: (String/None) Date of delivery, if applicable.
• item_category: (String) Category of the product (e.g., electronics, apparel, perishable).
• fulfillment_type: (String) Either first-party or marketplace_seller.
• shipping_region: (String) Geographic region for shipping (e.g., US-North, EU-France).
• order_status: (String) Current status (placed, shipped, delivered, returned).
• payment_method: (String) Method used for payment (Optional).

Getting Started

1. Clone the repository.
2. Install dependencies:

   pip install -r ecommerce_agent/requirements.txt

3. Set up Environment Variables: Create a .env file in the root directory with your Groq API key:

   GROQ_API_KEY=your_groq_api_key_here
   

4. Build the Index (Optional - Indices pre-built):

   python ecommerce_agent/src/ingest.py

5. Run Evaluation:

   python ecommerce_agent/evaluate.py

6. Launch Dashboard:

   python -m streamlit run ecommerce_agent/demo.py --server.port 8505

Deployment (Streamlit Cloud)

ZenithSupport AI is pre-configured for Streamlit Community Cloud:

1. Push to GitHub: Follow the instructions in the Walkthrough.
2. Deploy: Connect your repository to Streamlit Cloud.
3. Secrets: In the Streamlit Cloud dashboard, go to Settings > Secrets and add:

   GROQ_API_KEY = "your-gsk-key-here"

4. Main File: Point to ecommerce_agent/demo.py.

Evaluation Results

• Total Scenarios: 20
• Policy Volume: 27,065 words (across 14 documents)
• Citation Coverage: 100% (Doc + Section + Chunk ID)
• Core Accuracy: 100% (Grounded in retrieved sources via 4-agent validation)

Author

Karan Sharma