Boundary Detector

The 'Instructions' property allows you to provide custom, user-defined guidance to the LLM when it is tasked with identifying anchors for transaction boundaries. Use this property to clarify the type of anchors to look for, emphasize certain field labels, or instruct the model to ignore irrelevant or noisy content.

Usage and Impact

• Enter clear, concise instructions to improve anchor selection, especially for documents with complex or ambiguous layouts.
• Instructions are appended to the LLM prompt and can help resolve issues where default detection does not yield accurate results.
• If left blank, the LLM will rely solely on the job description and schema for anchor detection.

Example

Focus on field labels that appear at the start of each transaction, such as "Employee ID" or "Claim Number".
Ignore page headers and footers.
Only use regular expressions for numeric or date values.

Best Practices

• Use this property to address edge cases or document-specific quirks that may confuse the LLM.
• If anchor detection is unreliable, add explicit instructions to clarify what constitutes a valid anchor.

The 'Detection Window Size' property controls how many pages from the beginning of the document are analyzed for anchor detection. This allows you to focus detection on the most relevant portion of the document, improving both performance and accuracy.

• Set to a positive integer to limit detection to the first N pages.
• If the document contains fewer pages than this value, the entire document is analyzed.
• Useful for documents where transaction patterns are established early and repeated throughout.

Example

• Set to 5 to analyze only the first five pages for anchor patterns.
• Set to 1 to analyze only the first page.

Best Practices

• Use a smaller window for large documents with consistent structure to speed up detection.
• Use a larger window if the document has variable layouts or if anchors may appear later.
• This window should be large enough to expose at least 10 repeating transactions to the LLM during boundary detection.

Can be one of the following types:

The 'Document Quoting' property determines the quoting method used to provide document context to the LLM during both boundary detection and data extraction. Quoting methods control how text is formatted, delimited, and annotated in the prompt sent to the AI model.

Usage and Configuration

• Choose a quoting method that best preserves the structure and meaning of your document content.
• Some quoting methods add line numbers, delimiters, or special formatting to help the LLM understand document layout.
• Custom quoting methods can be used for specialized document types or extraction scenarios.

Impact

• The quoting method can significantly affect extraction accuracy, especially for documents with complex or ambiguous layouts.
• Consistent quoting improves the LLM's ability to align extracted data with the original document.

Example

• Use a quoting method that adds line numbers for multi-line transactions, or a delimiter-based method for tabular data.

For more information, see the documentation for Quoting Method.

The 'Anchor Count' property specifies how many distinct anchors the LLM should attempt to identify in the document. More anchors generally improve detection reliability, especially in documents with variable layouts or noisy content.

• Set to a higher value for complex documents with many fields or inconsistent structure.
• Set to a lower value for simple, highly regular documents.
• The allowed range is 1 to 16.

Example

• Set to 3 to require at least three anchors for boundary detection.

Best Practices

• Increase this value if boundaries are missed or if false positives occur.
• Decrease if the LLM struggles to find enough anchors in sparse documents.

The 'Minimum Score' property determines the threshold for how many detected anchors must match a line for it to be considered the start of a transaction. The score is calculated as the ratio of matched anchors to the total number of anchors.

• Set between 0 and 1 (e.g., 0.5 means at least half the anchors must match).
• Lower values increase sensitivity (more boundaries detected), higher values increase specificity (fewer, more confident boundaries).
• Adjust this value to balance between missing boundaries and over-segmentation.

Example

• Set to 0.5 to require at least half of the anchors to match for a boundary.
• Set to 1.0 to require all anchors to match.

Best Practices

• Lower the score if some valid boundaries are being missed.
• Raise the score if too many false boundaries are detected.

The 'Max Detection Attempts' property controls how many times the detector will retry anchor detection with alternative strategies or updated instructions if no valid boundaries are found on the first attempt.

• Each attempt may use different anchor sets or prompt variations to improve results.
• Previous unsuccessful anchor sets are provided to the LLM to help refine detection.
• The allowed range is 1 to 8.

Example

• Set to 3 to allow up to three detection attempts before giving up.

Best Practices

• Increase this value if your documents are highly variable or if anchor detection is unreliable.
• Decrease to speed up processing if detection is consistently successful on the first attempt.

Can be one of the following types:

The 'Fuzzy Matching' property allows you to configure fuzzy matching options for anchor detection, enabling the Boundary Detector to identify anchors even when there are minor OCR errors, typographical mistakes, or character substitutions in the document text. This is especially useful for documents with inconsistent quality, noisy scans, or variable data entry practices.

How Fuzzy Matching Works

• When enabled, anchor patterns (including static text and regular expressions) are matched using fuzzy logic rather than requiring exact character matches.
• The fuzzy matching engine evaluates the similarity between the anchor pattern and each line of document text, allowing for a configurable degree of difference (edit distance).
• Fuzzy matching can automatically correct or tolerate common OCR errors, such as 'O' vs. '0', 'B' vs. '8', or misplaced punctuation.

Configuration

The 'Fuzzy Matching' property exposes an FRX Options object, which includes:

• Match Mode:
  Controls how overlapping or ambiguous matches are resolved (e.g., best match, all matches).
• Weightings:
  Allows you to define or select a set of custom character-level edit costs using Fuzzy Match Weightings. This enables fine-tuned control over which character substitutions are considered more or less significant.

Usage and Impact

• Enable fuzzy matching for documents with poor OCR quality, frequent scanning artifacts, or known data entry errors.
• Adjust the match mode and weightings to optimize detection for your specific document type and error patterns.
• Disabling fuzzy matching (by leaving this property blank) will require exact matches for all anchors, which may reduce detection accuracy in noisy or variable documents.

Example

To allow for common OCR confusion between 'O' and '0', configure the weightings to set a low cost for this swap:

O0=0.25

This will increase the likelihood that anchors containing these characters are matched, even if the OCR result is imperfect.

Best Practices

• Use fuzzy matching selectively, as it may increase processing time for large or complex documents.
• Review anchor detection diagnostics to ensure that fuzzy matching is improving, not degrading, boundary accuracy.
• Combine fuzzy matching with well-chosen anchor patterns and minimum score thresholds for optimal results.

For more information, see the documentation for FRX Options and Fuzzy Match Weightings.