Six providers behind three actions, plus models you host yourself. The LLM integration has three actions. The bound connection's type decides the provider; nothing else in the workflow changes when you switch.
Output: text, aiResponse (same value), model, finishReason (stop, length, content_filter, other), usage (input and output tokens as reported by the provider), truncated.
No automatic retry: one attempt, 240-second timeout. A failed or timed-out generation is never re-sent, so it is never billed twice.
evaluate scores a text with the connection's model against one criterion and continues on the pass or the fail branch. It is advisory: never use it as an authorisation gate.
threshold 0–1 (default 0.7): pass when score ≥ threshold. Optional model overrides the connection's model. Output: criterion, score, passed, threshold, rationale, model.
The Model comparison template runs three generate → evaluate lanes on one question and merges the scores.
Every key is required; there are no anonymous model calls. The same credentials power Chat with Data and Designer Chat.
Administrators add open-weight models served by llama.cpp under Admin Center › Platform › Self-hosted models: display label, model id, base URL of the llama.cpp server, quantisation, context length and GPU layers (0 = CPU only).
This site needs JavaScript for the full experience. Every page above is reachable without it.