Project FORTHought A locally-hosted AI research configuration for Physics and STEM laboratories.

What This Is

This repository documents my working configuration for a locally-hosted AI research platform, built around Open WebUI, that I use daily in my Physics lab. It includes the custom tools, functions, optimizations, and MCP servers I've developed or assembled to support scientific workflows — literature review, spectroscopy, electron microscopy, X-ray diffraction, and data analysis.

This is not a product. It is a personal research setup that I maintain and iterate on. I upload my findings, configurations, and custom code here for documentation, reproducibility, and to share with my research group. The tools I build sit on top of existing open-source projects; my contribution is in how they are assembled, configured, and extended for real lab use.

Hardware:

The two machines communicate over Tailscale. The Docker server runs the full containerized stack, while the compute server handles GPU-intensive inference tasks (local LLMs, reranking, embeddings, document parsing).

Configuration note: All service URLs default to localhost. To run your own instance, copy config/.env.example to .env and fill in your values. If you use Cloudflare Tunnels or a reverse proxy, update the relevant environment variables.

Base Stack

The platform runs across two machines. The Docker server hosts the containerized stack, while the compute server handles GPU inference:

• LLM inference via LM Studio (local models on the compute server's dual R9700s) and cloud APIs (Gemini via Google research grant, OpenRouter as fallback)
• Code execution via a GPU-accelerated Jupyter kernel (ROCm, scientific Python stack)
• Document parsing via Docling (native PyTorch ROCm, with Qwen3-VL for figure descriptions)
• Vector search via Qdrant (hybrid BM25 + semantic)
• Tool orchestration via MetaMCP, which aggregates all my MCP servers behind a single HTTP endpoint

All services run on-premises. No data leaves the local network unless I explicitly use a cloud LLM.

What I've Configured and Built

RAG Pipeline Optimization

I tuned Open WebUI's RAG pipeline for scientific documents. Out-of-the-box defaults struggle with multi-column papers, equations, and dense tables. My current configuration:

... View full README on GitHub