AI providers
WME's AI features call an OpenAI-compatible chat-completions API through the backend. Provider keys never reach the browser.
| Feature | Needs |
|---|---|
| AI Logstash assistant | Any chat model; streaming responses |
| Workflow Assistant (Preview) | A model that supports tool / function calling |
Server default
Set in the backend .env:
AI_BASE_URL=https://your-ai-host.example.com/v1 # OpenAI-compatible base URL
AI_MODEL=google/gemma-4-26B
AI_API_KEY=... # required; use a placeholder for no-auth endpoints
AI_TIMEOUT_SECONDS=120
Users see this as System default in Settings → AI Assistant. If no server key is configured, users must choose their own provider.
WME requires a nonempty API key for both server and per-user providers. For a local endpoint that ignores authentication, use a placeholder such as local-no-auth.
Per-user providers
In Settings → AI Assistant → My own provider each user can set a base URL, model and API key (stored encrypted with CREDENTIALS_ENCRYPTION_KEY) and use Test connection. Presets are included for OpenAI, Groq, Together, OpenRouter and local Ollama / vLLM.
Self-hosted models
Any server that exposes /v1/chat/completions works — vLLM, Ollama (http://ollama:11434/v1), LM Studio, LiteLLM. Make sure the backend container can reach it, and for the Workflow Assistant pick a model trained for tool calling (for example recent Llama 3.x, Qwen 2.5+ or Gemma models served with tool-call support enabled).
Data handling
- The Logstash assistant sends the processor's configuration — including input/output resource parameters — and your messages to the model.
- The Workflow Assistant sends your messages and the catalogue data its tools return. Secrets entered in its secure forms are sent only to the backend.
- Choose a provider whose data policy fits the data in your catalogue.