Run ClassifAI features against on-prem LLMs and image models — Ollama for text and embeddings, Stable Diffusion for image generation.
How it works
ClassifAI is built around the principle that every feature should have a path that does not require sending site content to a third-party cloud. For language and embedding features that path is Ollama; for image generation it is a locally hosted Stable Diffusion installation through the AUTOMATIC1111 web UI, which exposes a compatible REST surface that ClassifAI can call.
Features with a local path
- Title generation, Excerpt generation, Key takeaways, Expand / Condense, Full-article drafting
- Alt text generation, Image tagging, OCR / text extraction
- Content tagging, Smart 404, Similar-term cleanup
- Image generation (via Stable Diffusion / AUTOMATIC1111)
Each of these features ships with a separate Ollama provider class for chat completion, vision, or embeddings models — so you can mix providers per feature, including an Ollama-only deployment that never leaves your VPC.
Trade-offs
The official docs are honest about the trade-offs: locally hosted models give you complete privacy and data control with no per-request cost, but performance is gated by your hardware and result quality varies by the specific model you have loaded. The most common production setup is a small dedicated GPU server inside the same VPC as the WordPress instance, so even the embeddings used for Smart 404 and Recommended Content never leave the customer’s network.
Providers
- Ollama (chat, vision, embeddings)
- Stable Diffusion via AUTOMATIC1111 (image generation)