FieldAlignMode

Specifies the method used to associate a value returned by the LLM with its location in the source document.

Remarks

The 'Field Align Mode' enumeration controls how Grooper links extracted field values from LLM output (typically in JSON) back to their original locations in the document. This association is essential for enabling users to navigate to, review, and highlight relevant content in the Data Grid and Data Viewer.

Each mode determines what information is requested from the LLM in the JSON schema, and how Grooper attempts to locate the value in the document. The choice of mode impacts review workflows, confidence scoring, and the user's ability to validate extracted data in context.

• None: No alignment is performed; only the value is requested and used as-is.
• Natural: The value is matched directly in the document text using normalization and type-specific matching.
• Quoted: The LLM is asked to return the exact text as it appears in the document (originalValue), which is used to locate the value in OCR text.
• Labeled: The LLM is asked to return a label (field name or header) from the document (label), and Grooper locates the value by finding the label and associating the nearest value.
• LabeledAndQuoted: Both a label and a quote are requested from the LLM, and both are used to disambiguate and precisely locate the value.
• Geometric: The LLM is asked to return explicit page number and bounding box coordinates (__pageNo, __bounds), which are used to directly extract the value from the document geometry.

Proper configuration of this setting improves the accuracy and usability of AI-driven extraction and review processes in Grooper. Choose the mode that best matches the structure of your documents and the nature of the field being extracted.

Can be one of the following values:

Name	Value		Description
None	0	►	No attempt is made to align the extracted value to a location in the document. Only the value is requested from the LLM in the JSON schema. The value is used as-is, with no attempt to match or highlight it in the document. No navigation or highlighting will be available in the Data Grid or Data Viewer. Use this mode when provenance is not important, or when alignment is not possible (e.g., for computed or synthetic fields). This mode provides the simplest and fastest extraction, but no review context.
Natural	1	►	Aligns the value by searching for an exact match in the document content. Only the value is requested from the LLM in the JSON schema. Grooper attempts to locate the value in the document text using normalization and type-specific matching (e.g., for numbers and dates, multiple formats are considered). If a unique match is found, the value is highlighted and navigable in the Data Grid and Data Viewer. If the value is ambiguous or appears multiple times, additional context may be needed. Use this mode for fields with unique, well-formed values such as IDs, numbers, or dates. Fast and reliable for simple fields, but may not work for repeated or ambiguous values.
Quoted	2	►	Aligns the value using a quoted passage from the document content. The JSON schema requests both the value and an originalValue property from the LLM, where originalValue is the exact text as it appears in the document. Grooper uses the quoted text to locate the value in the OCR text, even if the value itself is ambiguous or appears multiple times. This is especially useful for long text fields, descriptions, or values embedded in larger passages. If the LLM returns an inaccurate or incomplete quote, alignment may fail. Use this mode when the value may be ambiguous, or when you want to ensure the exact document text is captured and reviewable.
Labeled	3	►	Aligns the value by matching it to the closest label or field name in the document. The JSON schema requests both the value and a label property from the LLM, where label is the field name or header as it appears in the document. Grooper locates the label in the document and associates the value with the nearest occurrence. This is effective for forms, tables, or documents with clear field labels or headers. If the label is not unique or is missing, alignment may be ambiguous. Use this mode when values may be repeated but are always associated with a unique label in the document.
LabeledAndQuoted	4	►	Aligns the value using both a label and a quoted passage from the document. The JSON schema requests the value, a label, and an originalValue property from the LLM. Grooper first finds the label, then uses the quote to pinpoint the value's location. This provides the highest accuracy for complex documents, multi-line fields, or when both label and context are needed to disambiguate the value. Use this mode for fields where neither label nor quote alone is sufficient, or where maximum review accuracy is required. If either the label or quote is missing or inaccurate, alignment may fall back to the available information.
Geometric	5	►	Uses explicit page and location information provided by the LLM to align the value. This mode must be used in conjunction with the Layout Objects quoting method, which provides the location of each text segment. The JSON schema requests the value, __pageNo, and __bounds properties from the LLM. The LLM is asked to return the page number and bounding box coordinates (left, top, width, height) for the value. Grooper uses this information to directly extract and highlight the value from the document geometry. This is most useful when the LLM or extraction process can provide precise location data, such as in table extraction or when working with structured layouts. Use for advanced scenarios where spatial accuracy is critical, or when reviewing values in visually complex documents. If the LLM returns invalid or missing coordinates, alignment will fail.

Used By

AI Extract Field Options