AI - Data Extraction

Type identifier: ai:dataExtraction Category: AI Operations

Description

The AI - Data Extraction node uses large language models to extract structured data from unstructured text content. It can extract data matching a defined schema, optionally creating Records automatically upon extraction.

Input Handles

Output Handles

Configuration Options

Schema Configuration

The extraction schema can be defined in two ways:

Reference Mode:

Literal Mode:

Behaviour

1. Starts when its trigger input receives control; connecting only the data handles does not start extraction
2. Receives unstructured text content
3. Sends the content, the target object schema, and any Additional Instructions to the AI model

4. If Creation on Extraction is enabled:

   • Creates a Record for each extracted item
   • Creates the batch together: an insertion failure does not leave a partially created batch
   • Assigns each created Record an ID included in the extracted result
   • Does not create relationship links from extracted relationship values
   • Created Records are attributed to the run in the audit history, matching the Object - Create node
5. Returns extracted data and emits success only after any requested Record creation finishes

In the Flow run timeline, successful Record creation adds a message with the number of Records created and the target Schema ID. Select the extraction's completed attempt to see its creation summary: Schema ID, Schema version ID, created Record IDs, and total count. The summary lists up to 100 Record IDs and indicates when the list is truncated. The results output still contains all extracted Records, including their IDs.

AI billing events distinguish reserved credits (estimated capacity, not a charge), charged credits (final billed usage), and released reservations (no AI credits charged for that reservation). Actual usage can exceed the reservation. Select a billing event's Payload tab to inspect the reserved and charged amounts. Provider usage events show the model and estimated or actual token counts, including cache-token counts when available. Settlement based on an estimate rather than measured usage is marked Usage estimated. Input and output base credits are shown before the billing adjustment; reserved and charged totals include it. AI billing is separate from any Flow execution charge.

Extraction, validation, or Record creation failure emits the catalogued extraction_failed error through error, which can be routed by an Error-mode Control Switch. AI usage that has already been consumed remains chargeable even if validation or Record creation fails afterward. No response or a response without extracted action data is an error, not a successful empty result.

Extraction Modes

Examples

Extract Contact Information

Configuration:

• Extract Single: false
• Schema (reference): "Contact" Schema with name, email, phone

Input:

Please reach out to John Smith at john@example.com or call 555-1234.
You can also contact Sarah at sarah@company.org.

Output:

[
  { "name": "John Smith", "email": "john@example.com", "phone": "555-1234" },
  { "name": "Sarah", "email": "sarah@company.org", "phone": null }
]

Extract With Auto-Creation

Configuration:

• Extract Single: false
• Schema ID: "product-mention-schema"
• Creation on Extraction: true

Use case: Extract product mentions from reviews and automatically create Records.

Custom Extraction Instructions

Configuration:

• Additional Instructions:

  Focus on extracting only confirmed information.
  Mark uncertain values as null.
  Use ISO date format for any dates.