Extractor nodes¶
Extractor nodes read text and pull out structured data. Most extractors use an AI model (see AI models) to identify, filter or classify entities. They also capture the exact passage each detail came from so results can be verified later.
| Node | What it finds |
|---|---|
| Place extract | Locations mentioned in the text |
| Person extract | People |
| Organization extract | Companies, agencies, and other organizations |
| Custom extract | Capture types you define yourself, with your own fields |
Extracted people, places, and organizations can optionally flow into your Stylebook when the flow saves its results. See Canonicalization.
Grounded extraction¶
Entity extractors do more than return a name. They also identify the passage or passages that support the result. That evidence powers highlighted review in processed items and remains attached to mentions after the article is saved.
For example, Person Extract can distinguish a quoted person from someone merely named in the story, while Place Extract can preserve geographic hints that help the later Geocode step distinguish two places with the same name.
People, organizations, and places¶
- Person Extract can capture names, titles, affiliations, public-figure status, mention roles, and quote evidence.
- Organization Extract identifies companies, agencies, institutions, and other organizations, along with their role in the story.
- Place Extract identifies locations and their type, structured address or jurisdiction clues, and context useful for geocoding.
Each extractor has its own prompt and structured output format. The defaults encode Backfield's normal editorial distinctions; changing them can alter what downstream review and matching receive.
Custom Extract¶
Custom Extract is for structured information that does not belong in the shared person, organization, or location catalogs. A project might define fields for public meetings, election results, restaurant inspections, events, or recipes.
Custom records remain tied to the article and appear on its Custom review tab. Define fields that an editor can understand and verify, and include evidence-bearing fields where the record depends on wording in the source.
Model choice and prompts¶
Model-backed extractors use the project's approved AI models. Their panels expose the model and prompt used by future runs. A prompt change affects new executions only; past runs keep the output generated by their saved flow snapshot.
Extraction is a proposal, not a guarantee. Review important output against the source article, especially after changing a prompt, model, or output format.