AI Table Reader

Choose the model that best fits your extraction needs and available resources. The selected model determines the language understanding, accuracy, and cost of the extraction operation. Common choices include gpt-3.5-turbo for general-purpose extraction and gpt-4o for more complex or nuanced documents.

Guidance

• Use newer or more capable models for documents with complex structure or ambiguous content.
• Consider cost and performance trade-offs when selecting a model.
• The list of available models is determined by your LLM Connector configuration.

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.

Use this property to limit extraction to a subset of columns. This reduces prompt complexity and focuses the LLM on only the data you need.

• If left empty, all visible, non-computed descendant Data Columns are included.
• Add specific Data Columns to target extraction and improve performance.
• Useful for incremental extraction or when only certain data is needed from a document.

Example:
To extract only the "Invoice Date" and "Total Amount" fields, add those fields to this list.

Provide additional guidance or requirements for the LLM to follow during extraction. Instructions are appended to the prompt and can clarify expectations, formatting, or special rules.

• Use clear, concise language to direct the LLM.
• Examples: "Extract only dates in MM/DD/YYYY format." or "Ignore handwritten notes."
• Leave blank for standard extraction behavior.

Guidance on Prompt Engineering

Well-crafted instructions can significantly improve extraction accuracy, especially for documents with ambiguous or inconsistent layouts. Use this property to specify formatting requirements, business rules, or to clarify how the LLM should interpret certain document features.

• Be explicit about required formats (e.g., "Return all phone numbers in the format (###) ###-####").
• Instruct the LLM to ignore irrelevant content or to prioritize certain fields.
• For complex documents, break down instructions into bullet points or numbered steps.

Troubleshooting

If extraction results are incomplete or inconsistent, try refining your instructions. Test with different phrasings or add clarifying examples to guide the LLM toward the desired output.

Can be one of the following types:

The Document Quoting property determines how relevant portions of the document are selected and presented to the large language model (LLM) during table extraction. Quoting methods are essential for controlling the context, structure, and efficiency of the prompt sent to the AI.

Purpose and Usage

• Selects the scope of content to include in the prompt (e.g., entire region, specific fields, labeled sections).
• Applies optional preprocessing or formatting (such as tabular layout, JSON, or region marking) to help the LLM interpret complex tables.
• Reduces noise and prompt size by focusing only on the most relevant content for extraction.
• Adapts to different document layouts and extraction scenarios by allowing custom or built-in quoting strategies.

Configuration Guidance

• If left blank, the entire input region is included by default.
• Choose a quoting method that matches the structure and complexity of your tables (e.g., region-based, labeled, or layout-aware).
• For documents with ambiguous or variable layouts, targeted quoting can improve extraction accuracy and reduce LLM token usage.
• Custom quoting methods can be implemented for specialized scenarios, such as preprocessing, normalization, or advanced formatting.

Can be one of the following values:

The 'Row Align Mode' enumeration controls how Grooper requests and uses alignment information for table rows from the LLM. Each mode determines what properties are included in the JSON schema, what the LLM is asked to return, and how Grooper splits or locates table rows in the document.

• None: No alignment is performed; only the table data is requested and used as-is.
• Auto: Grooper infers row boundaries and alignment based on the positions of extracted cell values in the LLM response.
• Geometric: The LLM is asked to return explicit page number and bounding box coordinates for each row, which are used to directly extract the region.

Proper configuration of this setting improves the usability of AI-captured data for human review purposes.

Can be one of the following types:

The Multiline Row Settings class controls how Tabular Layout detects and assembles table rows that span multiple text lines. In many real-world documents, a single logical table row may be split across several lines due to text wrapping, stacked header layouts, or the inclusion of free-form comments and notes.

By configuring these settings, you can ensure that all relevant content is captured for each row, even when it is not confined to a single line.

How Multi-Line Row Detection Works

When enabled, multi-line row detection extends each detected table row to include additional lines that are determined to be part of the same logical row. This is especially useful for:

• Text Wrapping: Cell values that wrap onto subsequent lines.
• Stacked Layouts: Vertically stacked header cells with corresponding stacked values in each row.
• Free-Form Content: Comments, notes, or other information that appears below the main row data.

The detection process uses configurable limits for the number of lines, vertical spacing, and leading lines, and can be further tuned for page-wrapped or stacked layouts.

Configuration Guidance

• Use 'Maximum Lines Per Row' to limit how many lines can be grouped into a single row.
• Set 'Maximum Leading Lines' to allow for lines that precede the main row content, such as descriptions or comments.
• Adjust 'Maximum Line Spacing' to control how much vertical space is allowed between lines in a row.
• Enable 'Detect Page Wrap' to support rows that continue across page breaks.
• Enable 'Detect Stacked Layout' for tables with vertically stacked header and value cells.

Example Scenarios

• Invoice Line Items: Capture item descriptions that wrap onto multiple lines below the main row.
• Stacked Headers: Extract values from rows where each value is stacked vertically under its header.
• Free-Form Notes: Include comments or notes that appear below each row as part of the row's data.

Best Practices

• Test with a variety of sample documents to ensure multi-line detection is robust and does not over- or under-capture content.
• Use diagnostic output to review which lines are grouped into each row.
• Adjust limits and enable/disable features as needed to match your document layout.

For more information, see the documentation for Tabular Layout, Table Row Detector, and Data Table.

When enabled, the LLM is instructed to return data in a strict JSON format matching the extraction schema. This improves reliability and parsing accuracy, especially for complex or nested data.

• Supported by GPT-4o, GPT-4 Turbo, and recent GPT-3.5 Turbo models.
• When disabled, a best-effort parsing approach is used, which may be less reliable.
• Enable this option whenever possible for the most robust extraction results.