Developer shares five hard lessons building a GPU-free, on-premise RAG system for public sector

A developer working on public sector health information systems built a fully on-premise retrieval-augmented generation (RAG) stack that runs on CPU-only Windows Server hardware without Docker or cloud access. The open-source system uses ASP.NET Core 9, Ollama, Qdrant, Mistral 7B, and a Python ingest pipeline to process and query PDF, Word, and Excel documents locally. The project was motivated by strict institutional data-residency requirements common in healthcare and government environments, where modern AI infrastructure assumptions rarely apply. During deployment, the developer encountered five significant issues not covered in standard RAG tutorials, including gRPC incompatibility with HTTP/1.1-only corporate networks, which required bypassing the official Qdrant .NET SDK in favour of direct REST API calls. The field report is intended to help others facing similar legacy enterprise constraints where GPU access and cloud connectivity cannot be taken for granted.