FileAI

Inspiration

As a Computer Science student, I often found myself hunting down various files in my Downloads folder—homework, personal projects, resumes, cover letters, and images. My file organization was a mess, creating a sea of confusion and wasted time whenever I tried to search for or organize files. To overcome this, my team and I came up with FileAI.

What it does

FileAI uses semantic search to evaluate the contents of a file rather than relying solely on file names. This allows users to quickly locate files based on meaning, context, or related terms—even if the file is poorly named or buried in subfolders. It also provides a clean, interactive file explorer with chat-based assistance to preview, copy paths, and interact intelligently with files.

How we built it

Frontend: Built with React, Tailwind, and Cedar-OS components for the chat and file explorer UI.
Backend / Agent Layer: Powered by Cedar-OS and custom agents for file operations such as reading, searching, and organizing files.
Utilized Node.JS + TypeScript for dynamic backend for formatting API calls and server control.
AI Models: Semantic embeddings for search, paired with LLMs for natural interaction and contextual understanding.
HuggingFace Transformers + YOLO + FAISS CLIP for efficent semantic vector embedding and comparison.
Integration: Tauri was used to bridge system-level file operations with the web-based UI, ensuring a cross-platform experience.

Challenges we ran into

Ensuring semantic search reliability across diverse file formats (PDFs, images, text) was tricky. Solved by extensive preprocessing and format-specific embedding pipelines to normalize data.
Frontend-backend state integration with Cedar-OS and Mastra agents demanded careful synchronization. Used event-driven state mutations and clear API contracts to maintain smooth interaction and avoid stale UIs.
Avoiding UI clutter meant stripping debug elements conditionally and modularizing dev tools separately to keep production workflows clean.
Handling multi-step LLM response processing and routing through Cedar-OS’s structured pipeline ensured reliable UI updates and agent actions.
Challenges in cross-platform file access with Tauri were overcome by abstracting system calls and isolating platform-specific code paths for support for both Windows and iMacs

Accomplishments that we're proud of

Built a mini file explorer with AI assistance that blends traditional browsing with natural language search.
Successfully integrated semantic search into file operations, making searches context-aware.
Created a polished, responsive UI that adapts well to different themes and chat modes (floating, caption, side panel).
Learned to bridge AI tooling with real-world usability, focusing not just on AI but also the UX.
Solved a solution that many face, including myself, streamlining experiences with platforms with easier access to materials

What we learned

How to combine Cedar-OS agents with React components to create powerful interactive tools.
The importance of designing around usability—not just adding features, but making them intuitive.
Deepened our understanding of semantic embeddings and how to apply them in practical search scenarios.
Efficient techniques to embed various file formats into vector embeddings
Using local cacheing to optimize embedding comparison, similarity checks, and sorting
That debugging and cleaning up developer tools (like removing extra buttons) is just as critical as core features.

What's next for FileAI

Expanding file format support (e.g., code files, images with OCR, spreadsheets).
Adding automated organization features, where the AI can suggest or create folder structures and automatically sort new items.
Improving contextual previews so users can peek into files without opening them fully.
Enhancing security and permissions so FileAI can safely manage files without overreach.
Exploring integration with cloud storage providers (Google Drive, Dropbox, OneDrive).