Solar Flow -- A file organization master

Inspiration

We live in a world drowning in files. My own downloads folder had more than 1,500 files when I started this project. Existing tools organize files by name only, or rely on generic AI agents that quickly lose consistency in large folder structures. We wanted something smarter: a system that organizes files the way humans naturally do.

What it does

Solar Flow is an AI-powered file organization system that solves three core problems:

Deep File Structure Understanding
Problem: Current tools rely on file names and fail with complex folder structures
Our Solution: A tree-based approach with a custom database powered by embeddings
Innovation: A lightweight indexing agent that summarizes files and generates embeddings locally, running efficiently in the background
Intelligent Content Search
Problem: Users re-download files because they can’t recall names or keywords
Our Solution: Lightweight vector + keyword search, enabling discovery by both content and metadata
Benefit: Find files by what’s inside them, not just what they’re called
Human-like Organization Behavior
Problem: Tools don’t adapt to personal workflows
Our Solution: An organizer agent trained to mimic user behavior and suggest personalized structures
Result: Organization that feels intuitive and natural

How we built it

Architecture: Async pipeline (File → Parser → Summarization → Embeddings → Database → Search)
Backend: Python (Quart, SQLite + vector extensions)
Frontend: React 19 + Vite + Tailwind CSS with real-time progress tracking
AI Integration: Supports both local (Gemma3:4b, Ollama) and cloud models (OpenAI GPT-4, GPT-5, GPT OSS), plus Whisper for audio/video transcription
File Support: 30+ formats using Spotlight (macOS) with MarkItDown fallback
Search: Hybrid semantic + keyword scoring algorithm

Challenges we ran into

Time: As UW–Madison undergrads balancing research, startups, and internships, time was tight. AI tools like ChatGPT and Claude helped us accelerate brainstorming and coding.
Learning Curve: We were beginners in agent design and database architecture, so we spent significant time learning fundamentals.
Resources: No high-end GPUs. We optimized by using Mac hardware acceleration and carefully selecting efficient models.
Consistency: Ensuring agents behaved reliably required careful prompt engineering and rich context.
Compatibility: Supporting multiple embedding models forced us to design flexible, interconnected database tables.
Storage vs. Accuracy: Balancing speed, storage, and precision in search was a constant tradeoff.

Accomplishments that we're proud of

Working Prototype: A fully functional system built under time and resource constraints
Research Potential: A foundation to publish on LLM performance in file organization tasks
User Experience: Enhanced usability with AI-generated folder icons for visual navigation
Experimentation: Learned how to design agents, structure embeddings, and optimize pipelines in practice

What we learned

How to design consistent, context-aware AI agents
Building async pipelines with real-time progress tracking
Designing flexible vector databases that support multiple models
Integrating multi-modal AI for text, audio, and image content
Balancing performance, accuracy, and resource constraints

What's next for Solar Flow -- A file organization master

Native GUI: A SwiftUI-based macOS app
Model Fine-tuning: Optimizing Gemma models for indexing and OpenAI OSS models for organization
Codebase Rewrite: Improving maintainability and scalability
More File Types: Extending to specialized formats
Advanced Features: Folder watching, automated workflows, and batch operations
Academic Publication: Documenting our findings for the research community