Semantic

The 'Model' property specifies which language model will be used to generate vector embeddings for both the sample queries and the document content. The choice of model affects the quality and domain suitability of semantic matching.

• Select a model that is compatible with your document language and content type.
• Some models may require an API key or specific provider configuration.

For best results, test with several models if available, and review the documentation for your LLM provider.

The 'Queries' property allows you to specify one or more sample passages or clauses that represent the content you wish to find. Each query is compared to the document using semantic similarity, and the best-matching query determines the extracted content.

• Add multiple queries to improve detection when clause wording may vary.
• The highest-scoring query is used for extraction.

Example Queries: The parties agree to maintain confidentiality... All information disclosed shall remain confidential...

For each query, provide a representative example of the clause or passage to be detected.

The 'Preprocessing' property allows you to configure how the document text is normalized before chunking and semantic comparison. Preprocessing can include removing or replacing control characters, marking paragraphs, and other text cleanup operations.

• Use preprocessing to improve matching accuracy, especially for documents with inconsistent formatting.
• Adjust settings to match the structure and quirks of your source documents.

For more details, see the documentation for Text Preprocessor.

The 'Chunking' property controls how the document is divided into segments (chunks) for semantic comparison. Chunking parameters such as chunk size, overlap, and tokenization method can be adjusted to optimize detection for your document type.

• Smaller chunks increase granularity but may reduce context.
• Larger chunks provide more context but may miss short clauses.

Tune these settings based on the typical length and structure of the content you wish to extract.

The 'Clustering' property determines how top-scoring chunks are grouped and selected for extraction. Clustering helps combine adjacent or related chunks into a single result, improving the accuracy and completeness of extracted passages.

• Adjust clustering parameters to control sensitivity and grouping behavior.
• Use clustering to avoid fragmented or overlapping results.

For more information, see the documentation for Cluster Parameters.

Overview

The 'Name' property provides a concise, human-readable label for the quoting method. This value is used in Grooper’s user interface to help users identify, select, and configure quoting strategies. It is also injected into the prompt context sent to the large language model (LLM), where it serves as a section header or explicit label for the quoted content.

LLM Context Injection

• When the output of multiple quoting methods is combined (as in Multi Quote), the 'Name' is used to clearly distinguish each section of quoted content for the LLM. This helps the model understand that different representations or perspectives of the same document are being provided.
• In quoting methods such as Layout Objects, the 'Name' may be included in the prompt to clarify the structure or intent of the data, guiding the LLM’s interpretation and extraction logic.
• A descriptive and unambiguous name improves the LLM’s ability to reference, compare, and utilize the quoted data accurately, especially in complex or multi-view prompts.

Configuration and Usage

• Choose a name that reflects the quoting method’s purpose, formatting, or the type of content it produces.
• If left blank, the quoting method’s type name will be used as a default.
• Use consistent naming conventions to support clarity for both users and the LLM.

Example

• Claim Data
• Table Headers
• Layout Objects (JSON)

Best Practices

• Use names that are meaningful to both users and the LLM.
• When using Multi Quote, ensure each quoting method has a unique and descriptive name to maximize clarity in the combined prompt.
• Avoid generic names; be specific about the quoting method’s approach or output.

Overview

The 'Description' property provides a detailed, plain-language explanation of the content or data injected by this quoting method. It is intended to help the LLM understand the structure, semantics, and intended use of the quoted content, as well as to guide the model’s behavior when processing or extracting information from it.

LLM Context Injection

• The description is included directly in the prompt context sent to the LLM, typically as a prefix or annotation before the quoted data.
• In Multi Quote, each quoting method’s description is presented alongside its output, giving the LLM explicit instructions or context for interpreting each section.
• In Layout Objects, the description can clarify the meaning of structured layout data, JSON schemas, or other specialized formats, helping the LLM understand how to use the data for extraction or analysis.
• A well-crafted description can specify the format, highlight important features, explain relationships, or instruct the LLM on how to process the data (e.g., “This section contains a JSON array of line items, each with fields for description, quantity, and price. Extract each item as a separate record.”).

Configuration and Usage

• Write the description for the LLM, focusing on what the data is, how it is structured, and what the model should do with it.
• Include details about formatting, delimiters, field names, or any special conventions used in the quoted content.
• Use markdown formatting, examples, or sample data to illustrate the structure or intent.
• Update the description if the quoting method’s configuration or output changes.

Example

This section contains a table of invoice line items, formatted as plain text with tab-separated columns.
Each line represents a single item, with columns for description, quantity, and unit price.
Extract each line as a separate record and map the columns to their corresponding fields.

Best Practices

• Be explicit and unambiguous; avoid assumptions about the LLM’s prior knowledge.
• Tailor the description to the specific quoting method and output format.
• When using Multi Quote or Layout Objects, ensure each description is distinct and clearly describes the associated data.
• Use examples to illustrate complex or non-standard formats.

The 'Paragraph Padding' property controls how many paragraphs before and after a matching chunk are included in the final extracted result. This helps ensure that the full context of a clause or passage is captured, even if the match occurs in the middle of a paragraph.

• Set to null or leave empty to disable paragraph padding.
• Set to 0 to include only overlapping paragraphs.
• Set to a positive integer to include that many paragraphs before and after the match.

Use in combination with 'Word Padding' for fine-grained control over context inclusion.

The 'Word Padding' property allows you to include additional words before and after the matching chunk or paragraph, providing even more context for the extracted result.

• Set to 0 to disable word padding.
• Set to a positive integer to include that many words before and after the match.

This setting is most useful when precise control over the amount of surrounding text is required.