Roci AI

Inspiration

In today's information-rich world, valuable knowledge often lies buried within countless documents, inaccessible when you need it most. We built Roci AI to change that. Empowering people to unlock the full potential of their documents through intelligent retrieval. No more endless scrolling, no more forgotten insights. Just ask, and discover.

What It Does

Roci AI is an end-to-end AI document retrieval platform that uses contextual AI to deliver trustworthy answers from natural language queries, powered by your internal documentation. Simply upload your files, ask questions in plain English, and receive accurate, context-aware responses drawn directly from your own knowledge base. It's like having a research assistant who has read and memorized every document you've ever created.

How We Built It

Frontend: We adapted pluely (https://github.com/iamsrikanthnani/pluely) by Srikanth Nani (GPL 3.0 License, Open Source) to accelerate UI development, customizing it for our use case to create an intuitive "sit-in" conversational interface that feels natural and approachable.

Backend: Built entirely from scratch during the hackathon. We implemented a full RAG (Retrieval-Augmented Generation) system using Ollama 3.2, enabling users to search and query their documents using natural language. From document parsing to vector embeddings to intelligent response generation—every piece of the pipeline was crafted with care under the pressure of the clock.

Challenges We Ran Into

Document ingestion for RAG proved time-intensive, particularly when dealing with diverse file types. Each format brings its own quirks and edge cases. To stay focused and ship a working product, we made a strategic decision: optimize our current pipeline to handle .txt files first, ensuring a rock-solid foundation we can build upon. Another issue is that we had to develop the RAG + LLM model on one main laptop because of hardware limitations from 2 of our teammates. It slowed our development time a lot but we had to work around it to get it done

Accomplishments We're Proud Of

We're incredibly proud to have pushed out a fully functional RAG model in such a short timeframe. This idea has been living rent-free in our heads for a while now, and we seized this hackathon as the perfect opportunity to make a massive sprint from concept to reality. Seeing it come to life and actually work is deeply satisfying.

What We Learned

Building under pressure taught us the importance of scoping ruthlessly and shipping iteratively. We gained hands-on experience with chunking strategies, embedding pipelines, and the delicate balance between retrieval precision and speed. Most importantly, we learned that a focused MVP beats a sprawling prototype every time.