AI steps run inside a logic block, with full access to entity data and workspace isolation. The workhorse is ai_call.

ai_call

A single provider call for classification, extraction, summarization, generation, scoring, or translation. You give it context fields, an intent, and an output format; you get a typed result.

json 
{
  "type": "ai_call",
  "intent": "classify",
  "contextFields": [{ "label": "review", "value": "{{input.reviewText}}" }],
  "outputFormat": "category",
  "categories": ["positive", "negative", "neutral"]
}

The output shape follows outputFormat:

  • text → response (string).
  • json → a flat object with your declared outputFields.
  • score → score (0–100).
  • category → category matched against your categories.
  • boolean → decision (true/false).

Every result also reports tokensUsed. For json, list the fields you want extracted:

json 
{
  "intent": "extract",
  "contextFields": [{ "label": "email_body", "value": "{{input.emailText}}" }],
  "outputFormat": "json",
  "outputFields": [
    { "name": "sender_name",    "description": "full name of the sender" },
    { "name": "requested_date", "description": "meeting date in ISO format" },
    { "name": "topic",          "description": "one-line meeting topic" }
  ]
}

query_entity — RAG & memory

query_entity pulls entity records into the block so you can feed them to a model as grounded context, or load and save conversation state. Combine a search query with a generate step that's instructed to answer only from the retrieved documents.

text 
s1: query_entity  mode=search  entity=policy  search={{input.question}}  limit=3
s2: ai_call  intent=generate
             contextFields=[{ label: documents, value: {{s1.items}} },
                            { label: question,  value: {{input.question}} }]
             instructions="Answer only using the provided documents."
             outputFormat=text

ai_compare

Send the same prompt to several providers in parallel and get back an array with each provider's response, latency, and token count. Failures are isolated — one provider erroring doesn't abort the others.

json 
{
  "type": "ai_compare",
  "providers": ["anthropic", "openai", "gemini"],
  "userPrompt": "Summarize: {{input.articleText}}",
  "responseVariable": "comparison"
}

Persist the comparison

Follow an ai_compare with a create_entity step to store the ranked results — a simple multi-model product in two steps.

Agents & crews

Let a model call your logic blocks as tools.

AI integration

Configure the providers these steps use.