AI in flows

Flows have seven AI blocks. They use the models and vector fields you set up in AI and run on the server, so no key or rule has to reach the browser.

Block	What it does	Output
Ask a model (`ai.chat`)	Sends a prompt (or a whole conversation) to a chat model.	`text`, `message`, `finish_reason`, `usage`, and `json` when asked
Embed text (`ai.embed`)	Turns a text or a list of texts into vectors.	`vector`, `vectors`, `usage`
Search by meaning (`ai.search`)	Finds the records of a table closest to some text.	`data`, `count`
Answer from my data (`ai.answer`)	Searches, then has a chat model answer using only what it found.	`text`, `sources`, `usage`
Run an agent (`ai.agent`)	Hands a job to an agent and continues with its answer.	`run_id`, `thread_id`, `status`, `output`, `error`, `pending`, `cost`
Search a knowledge base (`ai.knowledge.search`)	Finds the passages of a knowledge base that best answer a question.	`results`
Add to a knowledge base (`ai.knowledge.add`)	Adds text to a knowledge base. The same `source` added again replaces what was there.	`document_id`, `status`

Ask a model

Setting	Meaning
`model`	A chat model from AI → Models.
`prompt`	What to ask. Templates work: `Summarise: {{ input.text }}`.
`system`	Instructions for the model.
`messages`	A whole conversation, instead of `prompt`.
`temperature`, `max_tokens`	Optional.
`json`	Turn on to expect JSON. The answer is parsed into `json`; a model that does not answer with JSON fails the step with `bad_output`.

Trigger

An HTTP trigger receives a support message.

Ask a model

prompt: Classify this message as billing, bug or other. Reply as JSON {"category": "..."}: {{ input.message }}, with JSON on.

Update record

data: { "category": "{{ steps.classify.output.json.category }}" }.

Answer from my data

The block most apps want: ask a question, get an answer grounded in your own records.

Setting	Meaning
`question`	The user’s question.
`resource`	The table to search. It needs a vector field that fills itself in.
`model`	The chat model that writes the answer.
`fields`	The fields to show the model, such as `title` and `body`.
`top_k`	How many records to retrieve. Default 4.
`max_distance`	Ignore records further away than this.
`system`	Replaces the default instructions: answer using only the context; if it is not there, say you do not know.
`filters`	Equality filters, for example only the current customer’s records.

It returns the answer and the records it used (sources), so you can show where it came from.

{ "block": "ai.answer",
  "config": { "question": "{{ input.question }}", "resource": "articles", "model": "chat",
              "fields": ["title", "body"], "top_k": 3, "filters": { "published": true } } }

Flows run with the developer’s authority: they can read any record. If a person’s question must only see that person’s data, pass filters that restrict it (for example { "owner_id": "{{ auth.user_id }}" }), or search with pawabase.from("…").similar(…) from the app, where the table’s access rule applies.

Search by meaning

ai.search takes resource and either text or vector, plus field, limit (default 5), metric, max_distance and filters. Vector fields are left out of the records it returns.

Errors

A model that is off, missing or of the wrong kind, or a provider that fails, fails the step with code ai_error and the provider’s reason. Add an on error branch to handle it, or give the model a fallback in AI → Models.

Run an agent

Setting	Meaning
`agent`	The agent’s name
`input`	What to ask it. Templates work.
`thread`	Carry on an earlier conversation: the `thread_id` of a previous run
`wait`	On (the default): the flow waits for the answer. Off: the flow carries on at once, and you start another flow from the `agent.run.finished` event.
`as_user`	Hold the agent’s tools to the access rules of the person who started this flow. Off (the default) treats the flow as trusted server code, so the agent’s tools are too.

If the agent needs a person’s approval, its status is waiting and pending lists what it is waiting on. See Approvals and limits.

Search and add to a knowledge base

ai.knowledge.search takes the knowledge base (base), a query and a limit (default 5), and outputs results, each with text, title, source and score. ai.knowledge.add takes the base, the text, and optionally a title, a source that identifies the document, and an owner_id that keeps it private to one person. Every block and every setting is listed in the AI block reference.

Getting started

Models

Agents

Knowledge bases

Connections and tools

Help

Ask a model

Answer from my data

Search by meaning

Errors

Run an agent

Search and add to a knowledge base

​Ask a model

​Answer from my data

​Search by meaning

​Errors

​Run an agent

​Search and add to a knowledge base

Ask a model

Answer from my data

Search by meaning

Errors

Run an agent

Search and add to a knowledge base