Creating a University Chatbot Using RAG

Introduction

University student portals host hundreds of pages of academic calendars, course lists, and exam schedules. Students struggle to find answers quickly. I built a RAG campus chatbot that indexes academic document files and answers queries accurately in real-time.

Background

University knowledge bases are updated regularly. Fine-tuning an LLM on calendars is slow and expensive. A RAG architecture allows the chatbot to retrieve source text directly from a vector database containing parsed documents and reference them in responses.

Implementation

The knowledge base is built on a vector index (Pinecone) using text-embedding models. When a user asks a question, the backend retrieves matching documents, builds a prompt, and streams the answer using the Gemini API.

Challenges

Deploying the campus chatbot revealed two primary challenges:

• Stale Data Conflicts: Old exam schedules conflicting with newer revisions in search results.
• Safety Concerns: Handling inappropriate user queries or off-topic questions.

Solutions

We solved this by establishing two key upgrades:

1. Metadata Filtering: Attaching versioning and semester tags to vectors to ensure the retriever filters for the active academic year.
2. Llama Guard Evaluation: Checking incoming queries and outgoing responses against safety policies before routing them.

Results

Adding semester metadata filters resolved data conflicts, and guardrails successfully filtered inappropriate inputs without increasing latency.

Conclusion

Building a successful RAG chatbot requires structured document processing. By filtering vectors using metadata and verifying messages with safety checks, developers build stable assistants for users.