A .NET 9 AI agent that keeps your data off the internet

What it does

AntSK is a full-stack AI knowledge base and agent platform built on .NET 9 and Blazor Server. It ingests documents—Word, PDF, Excel, Markdown, and more—into a vector store for RAG-style Q&A, exposes chat and plugin systems, and can run entirely on local models via Ollama, LlamaFactory, or llama.cpp GGUF weights. The stack also bundles Stable Diffusion for image generation, OCR, and an extensible plugin architecture supporting both .NET DLLs and HTTP APIs.

The interesting bit

The project’s scope is unusually comprehensive for the .NET ecosystem: it bundles document ingestion, RAG retrieval with BGE reranking, local Stable Diffusion, OCR, and a hot-pluggable DLL/API plugin system behind a single Blazor Server UI, targeting teams that want an all-in-one offline stack rather than a collection of separate services.

Key highlights

• Offline-first by design: runs local LLMs, embedding models, and Stable Diffusion without internet connectivity.
• Document pipeline: ingests Word, PDF, Excel, TXT, Markdown, JSON, and PPT into vector stores such as Qdrant, Redis, PostgreSQL/pgvector, or local disk.
• Plugin system supports hot-pluggable .NET DLLs, HTTP API plugins, and Semantic Kernel function calling.
• Multi-model backend: OpenAI, Azure OpenAI, Xunfei Spark, Alibaba Cloud, plus Ollama and LlamaFactory for local inference.
• Enterprise .NET stack: Blazor Server UI, SqlSugar ORM with PostgreSQL/SQLite/MySQL/SQL Server/Oracle support, and background task processing.

Caveats

• Model fine-tuning is listed as “planned” but not yet implemented.
• The public demos run on under-powered cloud servers that cannot host local models, so evaluating full offline behavior requires local deployment.
• Documentation and the primary README are Chinese-first; English support exists but its completeness is unclear from the provided sources.

Verdict

Teams running .NET infrastructure that must stay offline should evaluate this as a self-contained knowledge-base and agent host. Everyone else—especially those outside the Microsoft stack—will likely find lighter, language-agnostic alternatives.