AI nodes · Docs

Language-model steps on a Dataset › AI key: prompts, JSON, picking, writing, guarding.

Every AI call stops with the step: pausing, stopping or turning off a campaign aborts the request in flight (the provider may still bill what it already generated).

Generated from the node definitions in src/lib/flow/node/ai/ and the English labels (component.flow.node.ai). Every node also has the common settings.

Every AI node

Prompt, Extract JSON, Write, Summarize, Classify, Choose and Guardrail share three options:

May decline (Write, Extract JSON): the model may answer SKIP instead (nothing worth writing, it would repeat itself, it would break a rule), and the item leaves through declined. The posting and reply templates turn it on, so the AI never has to post.
Fallback AI key: used when the first key fails (its error is logged). Like AI key, it's one of the flow's AI key options ({{ option.ai }}), so a campaign picks the key.
Send the item's images: the item's image files (5 MB each) go with the question, for models that see images.
Memory key: the conversation under that key goes in before the question, and this turn is saved (Turns to remember, 10 by default). Use {{ item.sessionId }} in a chat.

AI agent — ai/agent
Choose — ai/choose
Embed — ai/embed
Knowledge — ai/knowledge
Classify — ai/classify
Guardrail — ai/guard
Extract JSON — ai/json
Prompt — ai/prompt
Summarize — ai/summarize
Write — ai/write
Answer — ai/answer
Memory — ai/memory
AI browse — ai/browse

AI agent

ai/agent · icon bot · plan feature ai.agent

An AI that can use tools before it answers. The engine runs its loop itself: the model answers, or calls tools, then reads their answers, up to Steps at most rounds. Answer as: text, JSON (shaped like Shape when set; one fix-up round when it isn't valid), or a list (one item per entry). Stream the answer sends its words as they come to the chat (or the streaming webhook) the run answers.

Field	Label	Kind	Default	Shown for
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key	dataset.ai
`system`	System message	prompt
`prompt`	Prompt	prompt	`"{{ item.message }}"`
`output`	Answer as	select: `text`, `json`, `list` (A list (one item each))	`"text"`
`schema`	Shape (JSON example)	json		output: json, list
`isStream`	Stream the answer (to the chat)	boolean	`false`	output: text
`tool`	Tools	list of { kind: select, name: text, description: textarea, flowId: flow, type: text, field: json, method: select, url: text, header: json, body: json, code: code, base: dataset.knowledge, topK: number }
`mcp`	MCP server	credential (mcp/server)
`mcpFilter`	Its tools	select: `all`, `only` (Only these), `except` (All but these)	`"all"`
`mcpTool`	Tool names (comma separated)	text		mcpFilter: only, except
`maxSteps`	Steps at most	number	`10`
`returnSteps`	Return the steps	boolean	`false`
`memory`	Memory key (empty: none)	text	`"{{ item.sessionId }}"`
`window`	Messages remembered (pairs)	number	`10`
`image`	Look at the item's images	boolean	`false`

A tool is one row of Tools:

Kind	What the model calls	How it runs
`flow`	A callable flow (its Flow input sample's keys are the arguments)	Walked inside the run like Run flow; what its Return sends back answers the model
`node`	Any node type (`type`, e.g. `data/http`), its `field` values as JSON	A step of that node in the run's browser or worker. `{{ ai.<name> }}` in its values are what the model fills, each a string argument. An effect node in a test run is simulated
`http`	`method`, `url`, `header`, `body` with `{{ ai.<name> }}` placeholders	From the engine, public addresses only; the answer is the status and up to 20,000 characters of the body
`code`	Your code, given the model's arguments as `input`	A step in the worker; what it returns answers the model
`knowledge`	`query`	A search of a (`base`, `topK` passages); the passages and their scores answer the model
`calculator`	`expression`	In the engine: numbers, `+ - * / % **` and parentheses
`think`	`thought`	Nothing happens: it gives the model room to plan

Plus every tool of the MCP server, when one is set (Streamable HTTP; its auth comes from the credential).

When the main AI key fails, the fallback key answers instead. Memory keeps the conversation under its key for the workspace, the last window pairs (test runs keep their own, test:<key>), so a later run with the same key picks up where it left off; it's kept 30 days after its last use. Each model call is logged (type ai, its tokens and cost count toward the run's AI totals) and each tool call is a log line with its input and output.

Output dots: next.

{ "text": "It's sunny and 31°C in Paris.", "steps": [{ "tool": "get_weather", "input": { "city": "Paris" }, "output": "[{\"weather\":\"sunny and 31C in Paris\"}]" }] }

json instead of text with Answer as JSON; steps only with Return the steps. Default timeout: 15 minutes.

Knowledge

ai/knowledge · icon library · plan feature ai.knowledge

Searches, adds to, updates or removes from a . Rerank fetches three times Results at most, then Cohere's reranker (a Cohere credential) or an AI key puts them in order and keeps the best. Update a document replaces the document of that id. Include metadata off drops each passage's metadata.

Field	Label	Kind	Default	Shown for
`base`	Knowledge base	dataset.knowledge
`operation`	Operation	select: `search`, `insert` (Add a document), `update` (Update a document), `delete` (Remove)	`"search"`
`query`	Search for	text		operation: search
`topK`	Results at most	number	`4`	operation: search
`minScore`	Score at least (0–1)	number	`0`	operation: search
`isMetadata`	Include metadata	boolean	`true`	operation: search
`rerank`	Rerank	select: `none` (No), `cohere` (With Cohere), `ai` (With an AI key)	`"none"`	operation: search
`cohere`	Cohere credential	credential (cohere/key)		operation: search; rerank: cohere
`ai`	AI key to rerank with	dataset.ai		operation: search; rerank: ai
`source`	Add from	select: `text`, `file` (A file)	`"text"`	operation: insert, update
`text`	Text	textarea		operation: insert, update; source: text
`file`	The item's file (key)	text	`"data"`	operation: insert, update; source: file
`name`	Document name	text		operation: insert, update
`isClear`	Empty the base first	boolean	`false`	operation: insert
`documentId`	Document id (replace or remove it)	text		operation: insert, update, delete
`metadata`	Metadata (JSON)	json		operation: insert, update
`filter`	Metadata filter (JSON)	json

Search sends one item per passage, { text, score, documentId, metadata }, best first; nothing found takes the empty dot. Insert sends { documentId, count } (its chunks); a file up to 10 MB in any of the supported formats. Delete sends { count } (documents removed): by documentId, or every document whose metadata matches filter.

Output dots: next, empty.

Embed

ai/embed · icon scan-text · plan feature ai.knowledge

Turns text into a vector with an OpenAI key's embedding model: { vector: [ … ], dimensions }. Fields: ai (OpenAI key), model (default text-embedding-3-small), dimensions (0: the model's own), text.

Field	Label	Kind	Default
`ai`	OpenAI key	dataset.ai
`model`	Embedding model	text	`"text-embedding-3-small"`
`dimensions`	Dimensions (0: the model's)	number	`0`
`text`	Text	textarea

Choose

ai/choose · icon list-checks

Picks the best items by your criteria

Fields

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key (when the first fails)	dataset.ai
`image`	Send the item's images	boolean	`false`
`memory`	Memory key (keeps the conversation)	text
`window`	Turns to remember	number	`10`
`list`	List	text
`criteria`	Criteria	prompt
`pick`	Pick	number	`5`
`extra`	Extra per pick	prompt

Output dots

next — Next
none — None

Output shape

Later nodes read it as {{ nodes.<id>.output… }}.

{
  "picks": [
    {
      "reason": "string"
    }
  ]
}

Default settings

timeout 2 minutes; the rest are the .

Example node

{
  "id": "choose",
  "type": "ai/choose",
  "name": "Choose",
  "field": {
    "pick": 5
  },
  "setting": {
    "retry": {
      "count": 0,
      "waitMs": 1000,
      "backoff": "fixed"
    },
    "error": { "mode": "stop", "fallback": null, "break": { "count": 3, "waitMs": 600000 } },
    "timeoutMs": 120000,
    "disabled": false
  },
  "pin": null
}

Classify

ai/classify · icon tags

Puts text into one of your categories

Fields

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key (when the first fails)	dataset.ai
`image`	Send the item's images	boolean	`false`
`memory`	Memory key (keeps the conversation)	text
`window`	Turns to remember	number	`10`
`input`	Input	textarea
`category`	Categories	list of { name: text, description: text }	`[{"name": "", "description": ""}]`

Output dots

next — Next

Output shape

Later nodes read it as {{ nodes.<id>.output… }}.

{
  "category": "string",
  "confidence": "number",
  "reason": "string"
}

Default settings

timeout 2 minutes; the rest are the .

Example node

{
  "id": "classify",
  "type": "ai/classify",
  "name": "Classify",
  "field": {
    "category": [
      {
        "name": "",
        "description": ""
      }
    ]
  },
  "setting": {
    "retry": {
      "count": 0,
      "waitMs": 1000,
      "backoff": "fixed"
    },
    "error": { "mode": "stop", "fallback": null, "break": { "count": 3, "waitMs": 600000 } },
    "timeoutMs": 120000,
    "disabled": false
  },
  "pin": null
}

Guardrail

ai/guard · icon shield-check

Checks text against your rules

Fields

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key (when the first fails)	dataset.ai
`image`	Send the item's images	boolean	`false`
`memory`	Memory key (keeps the conversation)	text
`window`	Turns to remember	number	`10`
`input`	Text to check	textarea
`rule`	Rules	prompt	`"No insults, hate, spam, links, personal data or false claims. Must make sense i`

Output dots

pass — Pass
fail — Fail

Output shape

Later nodes read it as {{ nodes.<id>.output… }}.

{
  "pass": "boolean",
  "violations": [
    "string"
  ],
  "input": "string"
}

Default settings

timeout 2 minutes; the rest are the .

Example node

{
  "id": "guard",
  "type": "ai/guard",
  "name": "Guardrail",
  "field": {
    "rule": "No insults, hate, spam, links, personal data or false claims. Must make sense in context."
  },
  "setting": {
    "retry": {
      "count": 0,
      "waitMs": 1000,
      "backoff": "fixed"
    },
    "error": { "mode": "stop", "fallback": null, "break": { "count": 3, "waitMs": 600000 } },
    "timeoutMs": 120000,
    "disabled": false
  },
  "pin": null
}

Extract JSON

ai/json · icon braces

Gets structured JSON from the AI

Fields

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key (when the first fails)	dataset.ai
`image`	Send the item's images	boolean	`false`
`memory`	Memory key (keeps the conversation)	text
`window`	Turns to remember	number	`10`
`input`	Input	textarea
`schema`	Schema	json	`{"title": "string"}`
`instructions`	Instructions	prompt

Output dots

next — Next
invalid — Invalid

Output shape

Later nodes read it as {{ nodes.<id>.output… }}.

"the item it got (unchanged), or what its code returns"

Default settings

timeout 2 minutes; the rest are the .

Example node

{
  "id": "json",
  "type": "ai/json",
  "name": "Extract JSON",
  "field": {
    "schema": {
      "title": "string"
    }
  },
  "setting": {
    "retry": {
      "count": 0,
      "waitMs": 1000,
      "backoff": "fixed"
    },
    "error": { "mode": "stop", "fallback": null, "break": { "count": 3, "waitMs": 600000 } },
    "timeoutMs": 120000,
    "disabled": false
  },
  "pin": null
}

Prompt

ai/prompt · icon sparkles

Asks the AI

Fields

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key (when the first fails)	dataset.ai
`image`	Send the item's images	boolean	`false`
`memory`	Memory key (keeps the conversation)	text
`window`	Turns to remember	number	`10`
`system`	System prompt	prompt
`prompt`	Prompt	prompt

Output dots

next — Next
error — Error

Output shape

Later nodes read it as {{ nodes.<id>.output… }}.

{
  "text": "string",
  "usage": {
    "input": "number",
    "output": "number",
    "cached": "number"
  }
}

Default settings

timeout 2 minutes; the rest are the .

Example node

{
  "id": "prompt",
  "type": "ai/prompt",
  "name": "Prompt",
  "field": {},
  "setting": {
    "retry": {
      "count": 0,
      "waitMs": 1000,
      "backoff": "fixed"
    },
    "error": { "mode": "stop", "fallback": null, "break": { "count": 3, "waitMs": 600000 } },
    "timeoutMs": 120000,
    "disabled": false
  },
  "pin": null
}

Summarize

ai/summarize · icon text-quote

Shortens text to what matters

Fields

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key (when the first fails)	dataset.ai
`image`	Send the item's images	boolean	`false`
`memory`	Memory key (keeps the conversation)	text
`window`	Turns to remember	number	`10`
`input`	Input	textarea
`length`	Length	select: `sentence` (One sentence), `short` (Short paragraph), `bullets` (Bullet points)	`"short"`

Output dots

next — Next

Output shape

Later nodes read it as {{ nodes.<id>.output… }}.

{
  "text": "string"
}

Default settings

timeout 2 minutes; the rest are the .

Example node

{
  "id": "summarize",
  "type": "ai/summarize",
  "name": "Summarize",
  "field": {
    "length": "short"
  },
  "setting": {
    "retry": {
      "count": 0,
      "waitMs": 1000,
      "backoff": "fixed"
    },
    "error": { "mode": "stop", "fallback": null, "break": { "count": 3, "waitMs": 600000 } },
    "timeoutMs": 120000,
    "disabled": false
  },
  "pin": null
}

Write

ai/write · icon pen-line

Writes a reply, post or message. The run log shows what it wrote (Reply: …), or Declined when the model answered SKIP.

Fields

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`fallback`	Fallback AI key (when the first fails)	dataset.ai
`image`	Send the item's images	boolean	`false`
`memory`	Memory key (keeps the conversation)	text
`window`	Turns to remember	number	`10`
`instructions`	What to write	prompt
`context`	Context	textarea
`tone`	Tone	select: `natural`, `casual`, `friendly`, `formal`, `witty`	`"natural"`
`length`	Length	select: `short`, `medium`, `long`	`"short"`
`language`	Language (optional)	text

Output dots

next — Next

Output shape

Later nodes read it as {{ nodes.<id>.output… }}.

{
  "text": "string"
}

Default settings

timeout 2 minutes; the rest are the .

Example node

{
  "id": "write",
  "type": "ai/write",
  "name": "Write",
  "field": {
    "tone": "natural",
    "length": "short"
  },
  "setting": {
    "retry": {
      "count": 0,
      "waitMs": 1000,
      "backoff": "fixed"
    },
    "error": { "mode": "stop", "fallback": null, "break": { "count": 3, "waitMs": 600000 } },
    "timeoutMs": 120000,
    "disabled": false
  },
  "pin": null
}

Answer

ai/answer · icon book-open-check · plan feature ai.knowledge

Answers a question from a : the nearest passages (Passages, 4) go to the model with the question and the instructions, and it answers with { text, source: [{ text, score, name }] }. Memory key (default {{ item.sessionId }}) keeps the conversation, its last window pairs (5).

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`base`	Knowledge base	dataset.knowledge
`question`	Question	prompt	`"{{ item.message }}"`
`system`	Instructions	prompt	`"Answer from the passages only. When they don't say, say you don't know."`
`topK`	Passages	number	`4`
`memory`	Memory key (empty: none)	text	`"{{ item.sessionId }}"`
`window`	Messages remembered (pairs)	number	`5`

Memory

ai/memory · icon brain

An AI conversation under a key, the same memory the AI agent and Answer use: Get its messages, Add a message (as the person or the assistant; the last window pairs kept), or Clear it. Output: the item with message: [{ role, content }].

Field	Label	Kind	Default	Shown for
`operation`	Operation	select: `get`, `insert` (Add a message), `delete` (Clear)	`"get"`
`key`	Memory key	text	`"{{ item.sessionId }}"`
`role`	As	select: `user` (The person), `assistant` (The assistant)	`"assistant"`	operation: insert
`text`	Message	textarea		operation: insert
`window`	Messages kept (pairs)	number	`10`

AI browse

ai/browse · icon mouse-pointer-click · plan feature ai.browse · effect

An AI that works the page toward a goal. Each step it reads the page (address, title, text, up to 80 numbered elements) and picks one action: click, type (not when Read only), go to a URL, scroll, done or fail; the browser does it. Allowed domains (comma separated) keep it on your sites: going anywhere else fails. Up to Steps at most (15). It ends through done ({ result, url, steps }) or failed (why). Each step is logged, its model call counted as AI.

Field	Label	Kind	Default
`ai`	AI key	dataset.ai
`goal`	Goal	prompt
`url`	Start at (URL)	text
`maxSteps`	Steps at most	number	`15`
`domain`	Allowed domains (comma separated; empty: any)	text
`isReadOnly`	Read only (never types)	boolean	`true`

Every AI node

Contents

AI agent

Knowledge

Embed

Choose

Fields

Output dots

Output shape

Default settings

Example node

Classify

Fields

Output dots

Output shape

Default settings

Example node

Guardrail

Fields

Output dots

Output shape

Default settings

Example node

Extract JSON

Fields

Output dots

Output shape

Default settings

Example node

Prompt

Fields

Output dots

Output shape

Default settings

Example node

Summarize

Fields

Output dots

Output shape

Default settings

Example node

Write

Fields

Output dots

Output shape

Default settings

Example node

Answer

Memory

AI browse