đź’ˇ Inspiration
We wanted to create a developer tool that simplifies working with GitHub repositories. Developers often deal with large codebases—our goal was to make understanding, managing, and navigating those projects easier with the help of AI.
⚙️ What it does
RepoFlow takes any public GitHub repo, clones it, generates a visual tree structure, extracts the README, and identifies all workspace-relevant files. Users can select a workspace (like a folder or module), and RepoFlow automatically generates vector-based context using a lightweight RAG model powered by an LLM. Developers can then chat with their code, getting insights, summaries, or suggestions instantly.
🛠️ How we built it
Backend: Python & FastAPI LLM + RAG: Gemini 2.5 Flash + SentenceTransformers + FAISS for vector store Git Integration: gitpython for repo cloning and management Tree & Workspace UI: Lightweight frontend using React (or CLI fallback) Chunking: Smart file chunking via LLM-based context understanding
🧱 Challenges we ran into
Designing an efficient and consistent file chunking strategy for RAG Managing memory usage when processing larger repos Ensuring accurate context matching using lightweight models Handling different repo structures while keeping the UX simple
🏆 Accomplishments that we're proud of
End-to-end working pipeline from GitHub URL to AI-powered repo assistant Clean modular design allowing developers to build on top of our project Fast response time even with large files thanks to optimized chunking and embedding One-click workspace focus for smoother debugging and exploration
đź“š What we learned
Practical application of Retrieval-Augmented Generation (RAG) in dev tools Real-world limitations of lightweight LLMs and how to prompt around them Importance of UI/UX simplicity even for backend-heavy projects Managing LLM context windows effectively for code understanding tasks
🔮 What’s next for RepoFlow
Upload-to-chat: allow users to upload local projects and chat with them Inline code editing + auto-commit suggestions via AI Support for private repos using GitHub OAuth Multi-agent mode: separate bots for documentation, optimization, bug-finding VSCode extension for full developer workflow integration
Log in or sign up for Devpost to join the conversation.