Slug: llm-taxonomy-model Version: 1.0.0 Infer: Yes Event Aware: No

​

Overview

​

LLM Data Labeling Model

The Kodexa LLM Data Labeling model enables intelligent extraction of structured data from documents using AI-powered recognition based on a data definition (taxonomy). It combines pattern recognition with Large Language Models to accurately identify and extract information from documents.

​

How It Works

1. Data Definition-Based Extraction: The model uses your data definition (taxonomy) to identify what data to extract from documents
2. Intelligent Chunking: Documents are broken into manageable pieces based on your selected strategy
3. Content Classification: AI analyzes content to identify relevant sections for each data element
4. LLM-Powered Extraction: Advanced language models process chunks to extract structured data
5. Merging and Validation: Results from multiple chunks are intelligently combined and validated

​

Core Features

• Flexible Chunking Strategies: Multiple approaches for breaking documents into processable pieces
• Multi-Level Classification: Identifies document types and specific data regions
• Automatic Labeling: Tags document pages and regions based on extracted content
• Embedding Support: Optional vector-based similarity search for classification
• Structure Review: AI-based validation of extracted structure
• Intelligent Merging: Combines results from multiple document chunks

​

Options Configuration

Option	Description
taxonomy	The data definition to use for extraction, defining the structure of data to extract
label_document	When enabled, adds labels to the document based on extracted data
set_external_data	Sets external data properties based on extracted structure
apply_guidance	Uses pre-defined guidance to improve extraction accuracy
external_data_key	Key to use when setting external data

​

Process Flow

​

Feature Options

​

Chunking Strategies

Controls how documents are divided for processing:

• Whole Document: Process entire document (limited by context window)
• Page: Process one page at a time
• First n Pages: Process only the first n pages
• Records: Detect and process individual records, even across pages
• Classified Content: Process all content of a specific type together
• Page Classified Content: Process classified content by page
• Consecutive Classified Content: Group consecutive classified content
• Group Classified Content: Group classified content across pages

​

Classification Strategies

Controls how document sections are classified:

• None: No classification
• Data Element: Classify based on data element definition
• Data Element and Children: Include child elements in classification
• Feature: Use custom feature for classification

​

Classification Content

Determines what content is used for classification:

• Text: Use text content (default)
• Bounding Boxes: Use spatial layout information
• Images & Bounding Boxes: Use visual and spatial information
• Image: Use only image data
• Embeddings: Use vector embeddings for similarity matching

​

Example Usage

The LLM Data Labeling model is particularly useful for:

• Extracting structured data from semi-structured documents
• Building automated document processing pipelines
• Creating training datasets for machine learning
• Validating document content against expected structure
• Integrating document information with downstream systems

​

Advanced Features

• Structure Review: Validates extracted structure against expected schema
• AI-Assisted Merging: Uses AI to resolve conflicts when merging data from multiple chunks
• Line Fallback: Automatically falls back to line-level processing for complex cases
• Thinking Mode: Enables LLM reasoning traces for better extraction quality
• Custom Model Selection: Override default models for classification and extraction

---

​

Inference Options

The following options can be configured when using this model for inference:

Name	Label	Type	Description	Default	Required
taxonomy	Data Definition	taxonomy	The data definition to use for the model	-	No
label_document	Label Document	boolean	Label the document	True	No
set_external_data	Set External Data	boolean	Set the external data to the structure from the data classes	False	No
apply_guidance	Apply Guidance	boolean	Apply guidance, if found	False	No
external_data_key	External Data Key	string	N/A	-	No

---

​

Model Details

• Provider: Kodexa

---

​

Feature Configuration Options

These options allow for fine-tuning the model’s features during configuration:

​

Option Group 1

Name	Label	Type	Description	Default
enable_line_fallback	Enable Line Fallback	boolean	Fallback to line level labeling if multiple lines and unable to find content	False
raise_exception_on_fallback	Raise Exception on Fallback	boolean	Raise an exception if we fallback to line level labeling	False

​

Option Group 2

Name	Label	Type	Description	Default
-	N/A	article	N/A	-
embedded	Embedded	boolean	Treat as embedded	False
cardinality	Cardinality	string	The cardinality of the data element in the chunk	single
classificationStrategy	Classification Strategy	string	Should we chunk using the data element for classification	dataElement
classificationContent	Classification Content	string	What content should we use for classification	text
maxHits	Max Embedding Hits	number	The maximum number of hits to return from embeddings	5
includeExplanation	Include Explanation	boolean	Try to include an explanation of the classification in the response, useful for debugging	True
ignoreNonWords	Ignore Non-Words	boolean	Ignore non-word tokens when classifying	True
restrictClassification	Restrict Classification	boolean	Restrict to only this classification (no mixed classes allowed)	True
rerank	Rerank Classification Matches	boolean	Should we rerank the classification results	False
maxPagesFromRerank	Max Pages From Rerank	number	The maximum number of pages to return from the rerank	5
chunkingStrategy	Chunking Strategy	string	How should we chunk the document for the LLM	classifiedContent
nPages	Number of Pages	number	The number of pages to use for the chunking strategy	5
tagPage	Label Page	boolean	Label the page, if classified	True
labelDocument	Tag Document	boolean	Add a tag to the document if classified	True
promptStrategy	Prompt Strategy	string	Which prompt strategy should we use	layout
image_width	Image Width	number	The width of the image	350
skipExtraction	Skip Extraction	boolean	Should we skip extraction for this data element	False
includeImages	Include Images	boolean	Should we include images in the prompt even if the strategy doesn’t normally include them	False
enableThinkingMode	Enable Thinking Mode	boolean	Should the LLM use thinking mode (if it is available for the selected extraction model)?	False
overrideExtractionModel	Override Extraction Model	boolean	Should we override the extraction model	False
extractionModel	Extraction Model	cloudModel	Choose a model if you wish to override the extraction model	anthropic.claude-3-haiku-20240307-v1:0
enableStructureReview	Enable Structure Review	boolean	Should we enable the structure review	True
structureReview	Structure Review Model	cloudModel	Choose a model if you wish to review the structure	anthropic.claude-3-haiku-20240307-v1:0
merge	Merge	boolean	Merge the objects identified in the chunks	True
mergeWithAI	Merge with AI	boolean	Use AI to review chunks that are grouped and merge them into a single representation	False
mergeInstructions	Merge Instructions	string	Additional instructions for merging the results to be included in the merge prompt	-