This is just a preview. The main results are on the table, expect an interesting blogpost. It will take another 2–3 weeks to finalize and will cover the following topics:
- Brief history of several millions of foreign forced laborers in Germany from 1942 to 1945
- The goal of research in more than 26,000 pages of historic forced labor interviews with local AI (no cloud usage, no tokens to pay for, and data privacy compliance) and how Retrieval-Augmented Generation (RAG) made the AI analysis possible
- Architecture of the local AI setup (hardware and software)
- Why I used an AMD Ryzen AI Max+ 395 (Strix Halo architecture) based Framework Desktop with Linux, VirtualBox, LM Studio, Open WebUI RAG, RAGFlow, and Hermes AI Agent – No Ollama, No OpenClaw AI Agent
- Benchmark between the local Qwen-3.8 27B and Google Gemma-4 31B large language models, which one of the famous open weights LLMs performed better?
- Challenges and solutions and why its good to have 48+ GB VRAM, virtual images and 6-bit LLM quantization instead of 4-bit
- Interesting findings from the WWII forced-laborer Interviews
- Lessons learned and tips and tricks
- Summary and outlook
Summary: Beyond the AI hype — What local AI can actually do when Retrieval Augmented Generation (RAG) is applied to more than 26,000 pages of historical PDF documents.
