project1

Advanced Generative AI – Capstone Project 1

Customer Support Chat (RAG)

This project is a Retrieval-Augmented Generation (RAG) based customer support system built with Python and Streamlit.
The system allows users to ask questions about provided documents and receive answers grounded in those documents.
If the answer cannot be found, the system suggests creating a support ticket.

---

Features

• Web-based chat interface (Streamlit)
• Document-based question answering (RAG)
• Semantic search using vector embeddings (ChromaDB)
• Answers are grounded in documents with source file name and page number
• Conversation history is preserved during the session
• Support ticket creation via external issue tracking system (Trello API)
• LLM-based decision making (function calling) for ticket creation
• System is aware of the company context (customer support scenario)

---

Data Sources

• At least 3 documents are used as knowledge sources
• At least 2 documents are PDF files
• At least 1 PDF document contains more than 400 pages
• Documents are ingested and split into chunks with page-level metadata

Example document:

• light_and_heavy_vehicle_technology.pdf

---

Supported Queries (Examples)

The system successfully answers document-grounded questions, for example:

"What's oil is better?"

This query works correctly and returns:

• Relevant information from the document
• Source file name
• Page number (e.g. light_and_heavy_vehicle_technology.pdf, page 116)

The answer is strictly based on the content of the documents (no hallucinations).

---

When No Answer Is Found

If the system cannot find relevant information in the documents:

• The user is informed that no document-based answer is available
• The system suggests creating a support ticket
• The user can manually create a ticket via the UI

This behavior is intentional to avoid hallucinations and ensure reliable answers.

---

Architecture Overview

1. User question is submitted via Streamlit UI
2. Question is embedded using sentence-transformers/all-MiniLM-L6-v2
3. Relevant document chunks are retrieved from ChromaDB
4. If relevant chunks are found:
   • The system displays the answer with source and page citation
5. If no chunks are found:
   • The system suggests creating a support ticket
6. LLM function calling decides whether a ticket should be created
7. Ticket is created in an external system (Trello)

---

Tech Stack

• Python 3.x
• Streamlit
• ChromaDB (vector storage)
• Sentence Transformers
• PyMuPDF (PDF processing)
• OpenRouter API (LLM)
• Trello API (issue tracking)
• dotenv (environment configuration)

---

How to Run Locally

pip install -r requirements.txt
streamlit run app.py