Spatial Grounding

Assigns spatial location data to fields within a Data Field Container using LLM processing.

Remarks

The Spatial Grounding fill method enriches extracted data by adding precise location information which links each field back to a page number and location on the source document. This is useful for post-processing outputs from LLM-based extraction mechanisms such as AI Extract, which do not inherently capture location data.

How It Works

• Prematching: Attempts to pre-match field values to their locations using OCR and document content before invoking the LLM.
• LLM Processing: Presents the LLM with both the extracted data and location data from the document, instructing it to update the JSON structure with spatial properties (pageNo, bounds).
• Results Application: Updates each field with the location information returned by the LLM, ensuring that extracted values are mapped to their exact position on the page.
• Configuration:
  • The 'Document Quoting' property controls how location data is presented to the LLM.
  • The 'Included Elements' property restricts which Data Elements are processed for spatial grounding.
  • Extension options allow fine-tuning for fields, tables, and sections.

Use Cases

• Document Review: Facilitates user validation by showing the spatial origin of each extracted value, improving transparency and trust in automated extraction.
• Quality Assurance: Enables detection of extraction errors by comparing field values to their actual location in the document.
• Workflow Integration: Supports downstream processes that require location-aware data, such as redaction, annotation, or visual overlays.

Extension Options

Spatial Grounding supports per-element customization through extension objects based on Data Element Extension:

• Spatial Grounding Field Options:
  Allows you to exclude individual fields from parent bounds calculations using the ExcludeFromParent property.
  This is useful when certain fields should not influence the computed spatial bounds of their parent section or table row.

• Spatial Grounding Section Options:
  Enables computation of section bounds from descendant fields after spatial grounding via the ComputeBounds property.
  This helps define the visual extent of a section based on the spatial positions of its child fields.

• Spatial Grounding Table Options:
  Provides the ability to compute table row bounds from descendant fields after spatial grounding using the ComputeBounds property.
  This is useful for visually grouping table rows or for downstream processes that require row-level location data.

These extension options are configured as embedded objects on their respective Data Field, Data Section, or Data Table elements, inheriting from Data Element Extension. They allow granular control over how spatial grounding is applied and how bounds are calculated.

Diagnostics

• Writes summary information to the diagnostic log, including the number of fields successfully pre-matched to location data.
• Adds the full LLM response as a JSON artifact (Response Data.json) for review and troubleshooting.

Notes

• The accuracy of location assignment depends on the quality of OCR data and the clarity of the document's layout.
• For best results, configure quoting and included elements to focus the LLM on relevant content and fields.

Properties

Name	Type		Description
General
Generator	Data Generator	►	The Data Generator used to run LLM completions for spatial grounding. EditorObject Properties Editor ConverterExpandable Converter Specifies the Data Generator responsible for interacting with the large language model (LLM) during spatial grounding. This generator executes the prompt and returns the updated JSON structure containing location data for each field. Choose a generator that supports the required LLM model and output format for your workflow. The generator's configuration (model, parameters, structured output) directly affects the accuracy and format of location assignments. For best results, use a generator compatible with the quoting method and document type. Example: Assign a generator configured for OpenAI GPT-4 with structured output enabled to maximize spatial mapping accuracy.
Document Quoting	Quoting Method	►	A quoting method which presents the LLM with location data extracted from the document. EditorObject Properties Editor ConverterType Selector Can be one of the following types: Value Description DI Layout Injects layout and semantic data from Azure Document Intelligence layout analysis into the prompt. JSON File Injects content from a JSON file stored on a Batch Object into the prompt. Multi Quote Combines multiple quoting strategies to generate a comprehensive document quote for AI extraction. Labeled Region Selects content occurring after a header label. Layout Objects Provides a quoting method that supplies the AI model with a page-by-page map of all visible content and layout elements in a document, including text, lines, barcodes, and checkboxes. Data Values Supplies the AI with extracted document data in JSON format. Extracted Selects all or part of a document as a quote for the AI model, with optional preprocessing for the text content. Semantic Selects portions of the document which are semantically similar to a set of examples. Controls how location data is extracted from the document and included in the LLM prompt. The quoting method determines the format and scope of spatial information provided to the model. Use Layout Objects to inject bounding box and page number data for text, barcodes, lines, and checkboxes. Use JSON File to provide output from DI Analyze or VLM Analyze activities, which supply location data for vision-detected constructs. The choice of quoting method impacts both the accuracy of spatial assignments and the cost of LLM processing. For large or complex documents, restrict quoting to relevant content to improve focus and reduce token usage. Example: Select Layout Objects to ensure the LLM receives detailed geometric data for all text elements on the page.
Included Elements	Data Element[]	►	Specifies which Data Elements to include in the spatial grounding operation. EditorIncluded Descendants Editor ConverterCollection Converter Limits spatial grounding to a subset of fields, tables, or sections within the Data Field Container. Only the listed Data Elements will be processed for location assignment. If left empty, all visible, non-computed descendant Data Elements are included by default. Add specific fields, tables, or sections to focus spatial grounding and reduce prompt complexity. Useful for incremental review, targeted validation, or when only certain data requires location mapping. The selection here should align with the extraction scope defined in AI Extract for consistent results. Example: Include only the "Invoice Date" and "Total Amount" fields to spatially ground key financial data for review.
Name	String	►	A descriptive name for this data fill operation. Requiredtrue The 'Name' property identifies this fill method within the list of fill methods attached to a Data Model, Data Section, or Data Table. Use a clear, descriptive name to indicate the purpose or strategy of the fill method (e.g., "Run Child Extractors", "External Lookup", "Custom Calculation"). The name is displayed in configuration UIs and logs, and is used to distinguish between multiple fill methods on the same container. The name is also used by the Fill Descendants command to identify which fill methods to execute during extraction.
Options
Trigger	Boolean Expression	►	An optional Boolean expression determining whether this fill method should run automatically during extraction. ConverterConverter The 'Trigger' property allows you to specify a condition under which this fill method will execute. If left empty, the fill method always runs. Enter a Boolean expression referencing data values, variables, or context to control execution. This enables conditional extraction logic, such as running a fill method only for certain document types or when specific data is present. Example: @DocumentType == "Invoice" && @TotalAmount > 0

See Also

Data GeneratorDI LayoutJSON FileMulti QuoteLabeled RegionLayout ObjectsData ValuesExtractedSemanticData ColumnData FieldData ModelData SectionData TableBoolean Expression