# Open GENAI

An unofficial project that runs Japan's Digital Agency GENAI government AI entirely locally, replacing cloud dependencies with SAML authentication through Keycloak, an OpenAI-compatible backend, RAG through Qdrant, and local transcription and image generation.

Open GENAI runs Japan's Digital Agency open-source government AI, called GENAI, in a fully local environment using a local, OpenAI-compatible large language model API instead of the original's cloud dependencies. It is an unofficial, community fork, unaffiliated with and not endorsed by the Digital Agency; it repackages the agency's own MIT-licensed frontend for local operation and does not include the agency's original backend, which must be obtained separately if needed.

Authentication runs entirely locally, with the project's backend acting as the SAML service provider and Keycloak as the identity provider; large-language-model calls go through an OpenAI-compatible API, defaulting to Ollama's endpoint but switchable to vLLM, LM Studio or another OpenAI-compatible server; and cloud services for chat history and inference streaming are replaced by a local FastAPI backend. Later versions add retrieval-augmented generation through Qdrant, speech transcription through Whisper, image generation through Stable Diffusion, multi-tenant knowledge and app separation, and personal-information detection in attachments.

It runs on macOS with Apple Silicon or on Linux with an NVIDIA GPU, including hardware such as the NVIDIA DGX Spark, and deploys with Docker Compose, with a separate production configuration terminating TLS.

Website: https://github.com/hirokawaguchi/open-genai