AI - Pawabase

AI blocks use the models, vector fields and agents you set up under AI. See AI in flows and Agents. This page documents 7 blocks. For how flows work, see Flows and the flow editor. Values in settings can use templates such as {{ input.id }} and {{ steps.load.output.email }}; see Block reference overview.

Block	What it does
Run an agent	Hands a job to an agent and continues with its answer.
Answer from my data	Answers a question from your own records: searches by meaning, then has a model answer from what it found.
Ask a model	Asks a chat model and returns its answer.
Embed text	Turns text into vectors.
Add to a knowledge base	Adds text to a knowledge base. The same source added again replaces what was there.
Search a knowledge base	Finds the passages in a knowledge base that best answer a question.
Search by meaning	Finds the records closest in meaning to some text (or a vector).

Run an agent

Key: ai.agent Hands a job to an agent and continues with its answer. Continues to: next

Setting	Type	Required	Default	Notes
`agent`	text	yes
`input`	text	yes		What to ask it
`thread`	text			Carry on an earlier conversation
`wait`	yes / no		yes	Off, the flow carries on at once. Start another flow from the agent.run.finished event
`as_user`	yes / no		no	Hold its tools to the rules of the user who started this flow

Answer from my data

Key: ai.answer Answers a question from your own records: searches by meaning, then has a model answer from what it found. Continues to: next

Setting	Type	Required	Default	Notes
`question`	text	yes
`resource`	text	yes		The resource to search
`model`	text	yes		The chat model that writes the answer
`fields`	list of string	yes		Fields to show the model
`field`	text			The vector field, when the resource has several
`top_k`	whole number		`4`
`max_distance`	number
`system`	text			Replaces the default instructions
`filters`	any JSON

Ask a model

Key: ai.chat Asks a chat model and returns its answer. Continues to: next

Setting	Type	Required	Default	Notes
`model`	text	yes		A model alias from the AI gateway
`prompt`	text			What to ask. Or use messages
`system`	text			Instructions for the model
`messages`	any JSON			A full conversation, instead of a prompt
`temperature`	number
`max_tokens`	whole number
`json`	yes / no		no	Expect JSON back and return it as `json`

Embed text

Key: ai.embed Turns text into vectors. Continues to: next

Setting	Type	Required	Default	Notes
`model`	text	yes		An embedding model alias
`input`	any JSON	yes		A string, or a list of strings

Add to a knowledge base

Key: ai.knowledge.add Adds text to a knowledge base. The same source added again replaces what was there. Continues to: next

Setting	Type	Required	Notes
`base`	text	yes
`text`	text	yes
`title`	text
`source`	text		What identifies it, so adding it again replaces it
`owner_id`	text		Only this user’s searches will find it

Search a knowledge base

Key: ai.knowledge.search Finds the passages in a knowledge base that best answer a question. Continues to: next

Setting	Type	Required	Default
`base`	text	yes
`query`	text	yes
`limit`	whole number		`5`

Search by meaning

Key: ai.search Finds the records closest in meaning to some text (or a vector). Continues to: next

Setting	Type	Required	Default	Notes
`resource`	text	yes
`text`	text			What to look for. Or use vector
`vector`	any JSON
`field`	text			The vector field, when the resource has several
`limit`	whole number		`5`
`metric`	`cosine \| l2 \| ip`		`cosine`
`max_distance`	number
`filters`	any JSON			Equality filters, `{field: value}`

​Run an agent

​Answer from my data

​Ask a model

​Embed text

​Add to a knowledge base

​Search a knowledge base

​Search by meaning