AI Separate

Specifies which large language model (LLM) will be used to analyze page content and determine document boundaries. The value must match a model ID available from the configured LLM Connector at the root of the repository.

Configuration Guidance

• Choose a model that is compatible with your organization's requirements for accuracy, speed, and cost.
• The available model IDs are determined by your LLM provider (such as OpenAI or Azure).
• More advanced models (e.g., gpt-4) may provide better results for complex or ambiguous documents, but may also incur higher costs or longer processing times.

Impact

The selected model directly affects the quality and reliability of document separation and classification. If you encounter unexpected results, consider testing with a different model or adjusting the 'Parameters' property.

Can be one of the following types:

The Chat Parameters class defines a set of options for tuning the behavior of LLM completion requests in Grooper. These parameters allow users to control the randomness, diversity, and topicality of generated responses, enabling tailored outputs for a wide range of scenarios.

Typical usage involves adjusting one or more parameters to guide the model’s output toward the desired style or content. For example, increasing randomness for creative tasks, or reducing repetition for summarization.

Configuration Guidance

• Adjust either 'Temperature' or 'Top P' to control randomness and diversity, but not both simultaneously.
• Use 'Presence Penalty' and 'Frequency Penalty' together to reduce repetition and encourage broader topic coverage.

Usage in Grooper

Chat Parameters are used throughout Grooper to influence LLM completions for document processing, data extraction, and other AI-driven tasks.

Impact

Proper configuration of ChatParameters can significantly affect the quality, creativity, and reliability of LLM outputs. Experimentation is recommended to find the best settings for your specific use case.

Allows you to supply additional instructions or context to the LLM for each separation analysis. These instructions are inserted into the system prompt for every page window, guiding the model's behavior and decision-making.

Usage Guidance

• Use this property primarily to define separation rules—criteria for when a new document should begin.
• Instructions can reference both page content and, if 'Include Page Metadata' is enabled, metadata fields such as page size, barcode presence, or custom values.
• While instructions may also touch on classification, the main source of classification guidance should be the 'Description' property of each Document Type.
• In practice, separation and classification logic often overlap, and instructions can reinforce or clarify document type distinctions as needed.

Example

Separation Guidance:

- Do not consider the following page types to be the start of a new document:
  - A blank page.
  - A page which only contains an imprinter / endorsement line (a date with some numbers).
  - A page which only contains a mailing label (name and address).
- Do not start a new document for attachments or exhibits to the previous document.

Impact

The quality and specificity of your instructions can have a significant effect on separation accuracy. Test and refine as needed for best results. By leveraging metadata, you can create highly targeted, context-aware separation rules that go beyond simple content analysis.

Can be one of the following types:

Controls how the content from each page is formatted and presented to the LLM for analysis. The quoting method determines which parts of the page are included, how they are structured, and whether any preprocessing is applied.

Configuration Guidance

• Select a quoting method that best matches your document structure and separation needs.
• If left unset, the entire document content will be included for each page.
• Some quoting methods may focus on specific fields, regions, or layouts to provide more targeted context.

Impact

The choice of quoting method can affect both the accuracy of separation and the cost of LLM usage (due to token limits and processing time). Use more focused quoting methods to reduce noise and improve efficiency.

For more information, see the documentation for Quoting Method.

Defines the size of the page window sent to the LLM for context. The total window size is calculated as 2 × Window Extent + 1, centering the analysis on the target page.

Configuration Guidance

• Increase the value to provide more context for the LLM, which may improve accuracy for documents with subtle or complex boundaries.
• Decrease the value to reduce processing time and token usage, especially for large batches or simple documents.

Example

• If set to 2, the window will include two pages before and after the current page (total of 5 pages).

Impact

A larger window may help the model understand document flow, but can increase cost and processing time. Adjust as needed based on your document set and performance requirements.

Enables classification of documents as they are separated. When set, the LLM will be prompted to select the most appropriate Document Type for each new document boundary.

Configuration Guidance

• Add one or more Document Types, Content Models, or Content Categories to include their descendant types.
• Each Document Type should have a clear, descriptive 'Description' property. This description is the primary source of information the LLM uses to distinguish between document types during classification.
• The 'Description' should focus on the unique characteristics, structure, and content of the document type, such as key phrases, layout, or required fields.
• While the 'Instructions' property may reference classification logic, the 'Description' property of each Document Type should be the main source of classification guidance.
• In practice, separation and classification often overlap, and both the instructions and descriptions may reinforce each other.

Examples of Document Type Descriptions

• Invoice: A document issued by a vendor listing goods or services provided, with line items, totals, and payment terms.
• Packing Slip: A document included with a shipment, listing the items shipped but not including prices or payment details.
• Statement: A summary of account activity, typically including a list of transactions, balances, and account information.
• Shipping Label: A single-page document containing sender and recipient addresses, a barcode, and carrier information.

Impact

Including document types allows for automatic assignment of types during separation, streamlining downstream extraction and validation. Well-written descriptions improve classification accuracy and help the LLM make distinctions between similar document types. Omit this property if you only need to split documents without classification.

When enabled, additional metadata from each Batch Page—such as custom values, descriptions, and barcode data—is included in the prompt sent to the LLM. This can provide valuable context for separation and classification.

Configuration Guidance

• Enable this option if your separation logic depends on metadata not present in the main page content.
• Metadata may include page dimensions, descriptions, custom field values, and detected barcodes.

Metadata Fields Included

The following metadata fields are injected for each page:

• widthInInches: The logical width of the page in inches.
• heightInInches: The logical height of the page in inches.
• description: The description of the page, if set.
• values: A JSON object containing all custom values stored on the page, except for documentTypeId and isNewDocument.
• barcodes: An array of detected barcodes, each with:
  • type: The barcode symbology (e.g., Code128, QR).
  • value: The decoded value of the barcode.

Impact

Including metadata improves the model's ability to detect boundaries based on non-textual cues and expands the types of instructions you can provide in the 'Instructions' property. For example, you can instruct the model to consider page size, specific metadata values, or the presence and type of barcodes when making separation or classification decisions. This enables more sophisticated, context-aware prompts such as:

• "Start a new document if the page width is greater than 11 inches."
• "Classify as 'Shipping Label' if a QR barcode is present."
• "Only separate if the 'department' metadata value changes between pages."

By leveraging metadata in your instructions, you can address complex separation scenarios that depend on both content and structured page attributes, resulting in more accurate and flexible document processing.

When enabled, the LLM is prompted to return its response as a structured JSON object, validated against a provided schema. This ensures consistent, machine-readable output for downstream processing.

Configuration Guidance

• Enable for scenarios where reliable parsing and validation of LLM output is required.
• If disabled, the model may return output in a less structured or more conversational format.

Impact

Structured output is recommended for most production scenarios, as it reduces the risk of parsing errors and simplifies integration with other Grooper components.

When enabled, the LLM is asked to provide a brief explanation for its decision to start a new document (or not) at each page. The reason is output in the diagnostics and is useful for troubleshooting and refining separation logic.

Enabling this feature may also improve accuracy in some cases, as the model is encouraged to think through its reasoning process before making a decision.

Configuration Guidance

• Enable this option to gain insight into the model's reasoning and improve transparency.
• Review the provided reasons in diagnostic logs to identify patterns or potential issues.
• Test with and without this option to see if it improves separation accuracy for your document set.

Impact

Including reasons may slightly increase prompt size and processing time, but provides valuable context for understanding and improving separation results.

Overview

The Bind Only property allows a Separation Provider to execute its page analysis logic and identify potential document boundaries, but without making any changes to the Batch Folder or its Batch Pages. No pages are moved, deleted, or classified, and no new folders are created. Instead, the separation points are recorded for later review and action.

Purpose and Role

• Separation Preview:
  Enables users to preview how the current separation logic will affect a batch, without risk of data loss or unwanted changes.
• Review and QA Workflows:
  Supports workflows where separation points are reviewed and confirmed by a user before being applied, typically using the Separation View in Grooper Review.
• Safe Testing:
  Allows designers to test and validate new or updated separation rules on real data, ensuring accuracy before committing changes.

Usage Scenarios

• Batch Import:
  Run separation in BindOnly mode during import to mark suggested document breaks, then allow operators to review and finalize separation.
• Quality Assurance:
  Use BindOnly to audit how separation logic would perform on production batches, identifying potential issues before they impact downstream processes.
• Iterative Design:
  Experiment with different separation strategies, reviewing the results in the Separation View before applying them.

Best Practices

• Always use BindOnly when developing or tuning separation logic to avoid unintended changes to batch data.
• Combine with the Separation View to provide operators with a visual interface for reviewing and confirming separation points.
• Once separation logic is validated and approved, disable BindOnly to apply the changes to the batch.

Related Features

• Separation View: The user interface for reviewing and confirming separation points identified in BindOnly mode.
• Batch Folder, Batch Page: The objects affected by separation when BindOnly is disabled.
• Separation Provider: The base class for all separation logic in Grooper.