RAG System for Company Documents

I co-developed an AI chatbot for BRUGG GROUP AG that helps employees find answers in company policies and manuals. The project used Retrieval-Augmented Generation (RAG).

Note: Due to confidentiality agreements, screenshots and detailed implementation examples cannot be shared publicly.

The Challenge

BRUGG's internal documents are confidential, so we couldn't use public AI services or store data outside the company. Everything had to stay within BRUGG's protected Microsoft Azure environment. I signed an NDA and worked through university access granted specifically for this project.

How We Built It

Working Environment: Six students collaborated through a dedicated Azure resource group (COHBChatbot-RG) that contained all our services: storage, AI models, and web hosting. We managed code through a private GitHub repository connected to Azure for automated deployment.

The Technical Setup:

• Document Storage: Company PDFs stored in Azure Blob Storage
• Document Processing: Python scripts broke documents into searchable chunks using LlamaIndex
• AI Models: Azure OpenAI's GPT-3.5-turbo for generating answers and text-embedding-ada-002 for understanding document meaning
• User Interface: Streamlit web app where employees type questions and see answers
• Deployment: GitHub Actions automatically deployed updates to Azure App Service

Design Decisions

We used a "temperature" setting of 0.2 for the AI model. Lower temperatures (like 0) gave incomplete answers, while higher ones (0.5+) made the AI ignore the documents. 0.2 gave us the sweet spot: accurate answers grounded in real company information.

Testing & Results

We tested with three BRUGG employees who each asked 10 questions:

• Simple queries: 7/10 average rating
• Complex queries: 5/10 average rating
• User interface: 5/10 rating

These scores showed the system works for straightforward questions but needs improvement for complex, multi-part queries.