bREIIn

bREIIn Home: Chatbot
Dashboard: Documents organized by topic
Dashboard: Knowledge graph showing information relationships
Inbox: Stored information by format (links, notes, documents, or audio)
Inbox: Uploading new information to the application

💡 Our Motivation: From RAG to GraphRAG

We are a team of four friends who have fought together in previous projects, but this time we wanted to push our limits. We had a clear goal: to work with RAG (Retrieval-Augmented Generation). However, after diving into the latest research, we decided not to settle for the basics. We wanted to experiment with GraphRAG.

The opportunity to transform flat, isolated data into an interconnected graph structure was the technical challenge we were looking for. bREIIn was born from that curiosity: we didn't just want to "search" for information; we wanted our AI to "navigate" the relationships between our notes, documents, and transcriptions.

👥 Teamwork and the MVP Challenge

The division of labor was natural: we split forces between Frontend and Backend. Within the backend, we specialized: while some focused on the AI model architecture, others tackled data preprocessing, API integration, and deployment.

One of our biggest challenges wasn't just the code, but product management. Brilliant ideas were constantly popping up, like a browser extension to capture web content instantly! However, we had to be disciplined. We constantly reminded each other of the importance of the MVP (Minimum Viable Product) to ensure a solid delivery, leaving those "big ideas" as our roadmap for the future.

🛠️ How we built it: The Intelligence Pipeline

Multi-modal Ingestion: We designed connectors for Google Keep (overcoming security hurdles via Master Tokens), PDF loaders, and an audio transcription module (Whisper) so that no idea is lost, regardless of its format.
The Brain (GraphRAG): We used LLMs to extract entities and relationships. We don't just retrieve text; we map concepts in Neo4j, creating a network of data where every link has semantic meaning.
Browser Extension: We managed to implement a prototype extension that allows users to inject information directly into the system while browsing, breaking the barrier between the web and your digital brain.
Tri-Database Architecture: In just a matter of hours, we implemented and synchronized three different database paradigms: a Graph database (Neo4j) for complex relationships, a Vector database for semantic RAG search (Qdrant), and a NoSQL database for fast document and metadata storage (MongoDB). A massive infrastructure deployment achieved in record time!

🚧 Challenges: Managing the Data Flow

It wasn't all smooth sailing. The greatest difficulty was managing the data flow between GraphRAG and automatic categorization. Making the AI understand what is a "person," what is a "technology," and how they should connect in the graph without generating visual noise was a process of constant iteration and prompt refinement. Additionally, integrating Google authentication from WSL environments tested us until the very last minute.

📈 What we learned

We learned that a graph is much more than a database; it’s a way of understanding context. We went from reading about GraphRAG to building a functional one, managing the entire lifecycle of an AI application—from preprocessing to interactive visualization.