Nodes

A node is one step in an Agate flow. Each node has a focused job and declares what information it needs and what it produces. In the guided builder, you add nodes with + controls; Agate creates compatible connections automatically.

Every flow starts with an input step and ends with an output step. In between, you add steps that find and refine the details you care about.

The node families

Nodes are grouped by what they do:

Family	What it does	Examples
Inputs	Bring text into the flow	Paste text, supply JSON (one document or up to 20 uploaded files), or pull a batch of files from cloud storage
Extractors	Pull structured details out of the text	Places, people, organizations, and custom record types you define
Enrichment	Refine extracted details or assign article-level metadata	Geocoding places into map coordinates; adding Article Meta tags
Embedding	Prepare content for semantic ("meaning-based") search	Indexing article text and images
Outputs	Save or export the results	Save into your catalog, view as JSON, or write files back to cloud storage
Flow control	Shape how data moves through the flow	Splitting long documents for extraction; gathering multiple branches together

A typical flow

Most flows follow the same left-to-right shape: bring text in, extract details, optionally enrich them, then save the results.

For example, a flow that maps the places in your reporting might be:

Paste text → Find places → Geocode → Save to database

How data moves between nodes

The article supplied by the input remains available throughout the flow. Specialized outputs travel along their branch: Place Extract produces places that Geocode can enrich, while Person Extract produces people. A later Gather step can collect several branches for an output that needs them together.

Agate uses these requirements to limit the add-step menu. If a node is not offered at a particular point, its required upstream data is not available there yet.

Node panels

Selecting a node opens its panel. The available tabs depend on the node:

• Settings controls its normal behavior.
• Prompt shows configurable instructions for model-backed extraction.
• Models selects approved model roles where a step needs more than one.
• Info explains inputs, outputs, and other details.
• Output shows that node's result when viewing a completed run.

The names and choices shown in a panel belong to the node itself, but model and integration options come from the current project's settings.

How nodes use AI

Many extractors and enrichment steps use AI models to read text and identify details. You choose which model a node uses from a list your administrators have approved, and Backfield tracks the estimated cost of each run. See AI models.

Not every node uses AI. Inputs, outputs, Gather, and some deterministic enrichment work without a generative model. A model-backed node should still produce structured output and evidence that editors can inspect later.

In this section

• Inputs — getting text into a flow
• Extractors — finding places, people, organizations, topics, and custom records
• Enrichment — refining extracted details, such as geocoding
• Embedding — preparing content for semantic search
• Outputs — saving results and exporting
• Flow control — combining branches