Workflow Nodes

The workflow editor registers the nodes below, grouped by the four palette categories. The exact input/output fields and configuration options for each node are determined by the editor configuration and the node reference.

Types that are not in the palette: http_request and document_extractor have executors but no node component, so they only appear in definitions submitted through the API; iteration_exit / loop_exit are end markers inside an Iteration/Loop container with no standalone executor and must not be connected as independent nodes.

Principles for Choosing Nodes

• Use LLM when chat model reasoning is needed; use Decision when you must pick one of several options or ordered levels and know how confident the model is; use Agent when you need to reuse complete Agent behavior.
• Use Knowledge Base Retrieval when you need to answer based on a knowledge base, then map the results to an LLM or output.
• Use Tools when you need to call existing HTTP/MCP/code capabilities; do not duplicate stored credentials in workflows.
• Use Iteration to process each array item; use Loop to repeat until a condition is met.
• Use Sub-workflow to isolate reusable flows, configuring input/output mappings and failure strategies.

Decision Node

The Decision node is the only node that returns probabilities: it hands a state to a decision model, which answers a single question and returns the probability distribution over options/levels plus a confidence; the node then takes the matching branch. Nothing is generated, so the route never has to be parsed out of model text and cannot drift because the model added a sentence. Full configuration steps and a worked example: Orchestrate a Workflow → Decision Node.

How it differs from LLM and Question Classifier

Rule of thumb: when "which path" depends on natural-language meaning and you also need to know how sure the model is, use a Decision node; for pure text generation use LLM; when there is a small fixed set of categories and probabilities do not matter, the Question Classifier is lighter.

The three question types

noul returns no confidence, so the confidence threshold has no effect on it; switch to choice or score when you want "route to a human when unsure".

Branch handles, fallback, and threshold

• The text of an option (choice) or level (score) is the name of that output handle on the canvas; the fallback handle is default unless you rename it in Fallback branch.
• Editing option text remaps existing connections by position as long as the count is unchanged. Leftover connections after a count change are reported by the checklist as an outdated output handle and must be reconnected.
• Confidence threshold (0–1, may be left empty) reroutes the flow to the fallback branch only when the returned confidence is strictly below it; a confidence equal to the threshold still takes the model's branch. The threshold affects routing only: answer is always the option/level the model picked, while selected_handle is the branch actually taken.
• With no threshold set, only "no branch handle resolved" uses the fallback. A fallback branch with no downstream nodes ends the flow at this node.
• A decision node is retried up to 2 times by default (1 second initial delay, exponential backoff with jitter).

Outputs and failures

Output variables: answer (the selected option/level, or yes/no), selected_handle (the branch actually taken), usage (token usage), plus choice + confidence + probabilities, score + confidence + probabilities, or noul depending on the type.

When it fails: missing configuration or invalid options fail as validation_error before the model is called; an answer whose type, option set, or probability set does not match the configuration fails as decision_result_missing; an unauthorized or disabled model, an endpoint blocked by the allowlist, a timeout (default 60 seconds), or exhausted quota surfaces as the corresponding provider error.

Debugging a decision branch

• The checklist checks each item: model not selected, empty state, empty instructions, too few or too many options/levels, empty or duplicate options, and connections using an outdated handle.
• In the run trace, use selected_handle to see which branch ran, confidence to see whether the threshold rerouted it, and probabilities to decide between adjusting the threshold and rewriting the instructions.
• A decision node is one call with one result and no intermediate steps, so there are no breakpoints or single-step debugging — you debug it by running the workflow.

Node-level Failures

Each executed node records inputs, outputs, model, Tokens, duration, retry count, error type, and error message. HTTP, code, MCP, and media nodes should configure an acceptable timeout; avoid unconditional repeated calls to side-effect interfaces that cannot be retried.