Overview
The Document Data Extraction task is designed to surface only the values that are explicitly present in the linked source documents. The extraction logic is configured to return values verbatim from document content, meaning it should not introduce or generate values that do not exist in the source material. As such, all extracted values are expected to match the document text.
Expected Behavior
Values are extracted directly from the content of linked documents.
Extracted data reflects the wording found in the document.
Each extracted value is associated with a source document within the journey.
Why Unexpected Values May Appear
Although the system is designed for verbatim extraction, there are some scenarios where values may appear that are not visibly present in the document being reviewed:
Inference from Contextual Cues
For certain fields that map to a limited set of real-world values (e.g., Primary Regulator), the system may infer a value based on strong contextual signals such as jurisdiction, license references, or descriptions of regulated activities.
In this case, the value is a plausible inference, not a direct quote.
Value Sourced from Another Document in the Journey
Extraction runs across all eligible documents linked to the journey, not just the most recently uploaded document.
The extracted value may originate from a different document within the same journey.
Cross-Field Contamination
When multiple related fields are extracted at the same time, a value identified for one field may occasionally be assigned to another similar field.
This can happen when field definitions or labels are closely related or overlapping.
Summary
Document Data Extraction is intended to return only values present in source documents. However, in certain cases, values may appear due to contextual inference, extraction across multiple documents, or overlap between related fields.