Generative AI

Decision AI or Generative AI?

Kovca has two kinds of AI Nodes, and a workflow can combine both. Pick the one the step needs:

• Decision AI makes decisions from defined possibilities: a category, a priority, a score, yes or no. It answers in about a tenth of a second at a fixed price per evaluation. Use it for decisions, and route on its outputs with conditions and switches.
• Generative AI creates and interprets content from flexible instructions: a personalized reply, a summary, details pulled out of a document, a translation, or a written analysis. It returns text or structured fields, usually takes a few seconds, and is metered by tokens.

A typical workflow uses both: Decision AI classifies a support request, a condition routes it, and Generative AI drafts the reply only on the path that sends one. The Studio template Support auto-reply does exactly that, and the Studio assistant builds workflows the same way: decisions with Decision AI, writing with Generative AI.

Prebuilt Generative AI

Under AI Nodes → Generative AI in the node library, choose Browse and add the ones you need to the workflow's library:

• Generate text: Write emails, replies, messages, and descriptions.
• Summarize: Condense documents, conversations, and earlier results.
• Extract data: Pull fields like names, dates, and amounts out of text.
• Transform: Rewrite, translate, or reformat text.
• Reason & analyze: Explain risks, causes, or trade-offs that fixed options can't capture.

Drag one onto the canvas, then set it up in the inspector:

• Instructions: describe the result in plain words. Click a value under the text box to insert it, such as {{input.message}} or {{nodes.triage.output.category}}. The node reads only the values you insert.
• Returns: Text gives later steps nodes.<id>.output.text. Fields returns the fields you define (text, numbers, true/false, one-of values, lists, or objects), each available as nodes.<id>.output.<field>. Describe each field so the AI knows what goes there; a field the text doesn't contain is left out (and reads as null downstream), never guessed. Answers that don't match the fields fail the step.
• Model settings (optional): the speed level (Fast, Balanced, or Advanced), creativity, maximum length, and reasoning depth. Each node already picks a sensible level.

Custom Generative AI

In the same catalog, open Custom and choose Create Generative AI to package instructions you'll reuse: typed inputs, instructions that read them ({{input.<field>}}), and the text or fields it returns. Like Custom Decision AI, it appears in every workflow's library, workflows map their values into its inputs, and you test, version, and publish it from Edit node. A published workflow uses a published version and stays pinned to it.

Test either kind from the Test panel or the node editor: the output, speed level, tokens used, and time appear on the step.

Data and safety

• Values are sent to the model as data, separate from your instructions. Instructions that appear inside customer text are ignored.
• Models run through Kovca's servers; no credentials reach the browser. Kovca stores token counts and timings only, never the values sent or the answers.
• A Generative AI node never sends anything anywhere by itself. To email a reply or post it to a system, connect it to an HTTP action.

Runs

A Generative AI node's output is part of the workflow output like any other step. GET /v1/runs/{id} reports the tokens each one used:

{ "id": "reply", "status": "succeeded", "evaluation": false, "duration_ms": 1840, "tokens": { "input": 412, "output": 96 } }

Usage and limits

• Generative AI uses credits by the tokens each call reads and writes (thinking counts as output), at least 1 credit a call. Credits per 1M tokens, input / output: Fast 500 / 3,000, Balanced 1,500 / 7,500, Advanced 4,000 / 24,000. Before a step runs, Kovca holds the credits of its longest possible answer and then keeps only what it used. Steps are limited to 20 (Free) or 120 (Pay As You Go) a minute. Past your credits or the rate, the step fails with 429 USAGE_LIMIT_EXCEEDED or 429 RATE_LIMITED.
• Each step reads at most 60,000 characters of values, answers in at most 8,192 tokens, and stops after at most 60 seconds, leaving a few seconds of the workflow's 60-second limit so the answer can be saved. A workflow can have up to 10 Generative AI nodes.
• Generative AI nodes can't run once per list item. Pass the whole list instead; one step can work through every item.
• In Instant Workflow trials, Generative AI runs for real within the trial's own allowance, with shorter limits (20,000 characters in, 2,048 tokens out), up to 6 runs a minute however many Generative AI nodes a run contains, and a shared workflow can have up to 5. Checking examples while a workflow is built doesn't use that rate.
• Errors: 422 UNPROCESSABLE_CONTENT when the answer doesn't match the node's fields, is cut off at its maximum length, or the model declines; 504 TIMEOUT and 503 SERVICE_UNAVAILABLE when generation is slow or unavailable. See Errors and limits.