Chat example of the GGV DOST
The interface of the GGV DOST
Icon of the GGV DOST

GGV DOST — Data-Driven Onboarding Support Tool

Inspiration

Guru Ghasidas Vishwavidyalaya’s website contains a wealth of academic and administrative information, but it is spread across hundreds of webpages, notices, and PDFs. As students, we often found ourselves manually searching through multiple links just to answer simple questions like deadlines, procedures, or eligibility criteria.

This friction inspired us to ask a simple question:

What if the entire university website could be queried like a conversation?

That idea became GGV DOST — an AI-powered conversational assistant designed to simplify information access for students and staff.

What We Built

GGV DOST is a smart chatbot that centralizes official university information and delivers accurate, context-aware answers using semantic search and Retrieval-Augmented Generation (RAG).

Instead of keyword-based searching, users can ask questions naturally and receive direct, trustworthy responses grounded in official data sources.

How We Built It

The project was designed as a production-style AI pipeline, even though this is a hackathon MVP.

1. Data Collection & Processing

Scraped official university webpages
Cleaned and converted raw HTML into structured text
Applied semantic chunking to store information in context-aware units

2. Semantic Search

Converted documents and queries into vector embeddings using Gemini embeddings
Stored embeddings in a MongoDB vector database
Retrieved the most relevant chunks using Top-K vector similarity search

Mathematically, similarity is computed using cosine similarity:

[ \text{similarity}(q, d) = \frac{q \cdot d}{|q| |d|} ]

where ( q ) is the query embedding and ( d ) is a document embedding.

3. Answer Generation (RAG)

Retrieved content is injected into prompts
A Gemini LLM generates answers strictly grounded in retrieved data
This minimizes hallucinations and improves trustworthiness

4. System Architecture

Frontend: React-based conversational UI
Backend: FastAPI for request handling and AI orchestration
Deployment: Dockerized services running on Google Cloud Run

What We Learned

How semantic search is fundamentally different from keyword search
Designing and implementing a Retrieval-Augmented Generation (RAG) pipeline
Building an end-to-end AI system: from data ingestion to cloud deployment
The importance of grounding LLM responses in verified data
How to balance technical depth with usability in real-world systems

Challenges We Faced

Data inconsistency: University webpages vary widely in structure and formatting
Chunking strategy: Finding the right balance between context size and retrieval accuracy
Time constraints: Building a full pipeline within a hackathon timeline
Scope control: Prioritizing core functionality over polish and extra features

Despite these challenges, we successfully built a functional prototype with the core system working end to end.

What’s Next

Expand data coverage to include PDFs and notices
Add multilingual support
Improve response explainability with source citations
Deploy the system campus-wide and adapt it for other universities

Conclusion

GGV DOST demonstrates how AI can transform fragmented institutional information into a single, intelligent conversational experience. While this is an MVP, the foundation is scalable, extensible, and impactful — making information access faster, simpler, and smarter.