Annotated Document Processing with Taxonomy-Augmented LLM Prompts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generative artificial intelligence (AI) applications using large language models (LLMs) face challenges in interacting with large taxonomies that exceed token limits, managing hierarchical structures, and incorporating user-specific data effectively.
Innovation Solution
The implementation of Taxonomy Augmented Generation (TAG) tactics, including attribute trimming, hierarchical diving, hierarchy flattening, singular item focusing, and borrowing alignment, to frame custom taxonomies within LLM prompts, ensuring they fit within token constraints and align with hierarchical structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If large taxonomies are incorporated into LLM prompts, then the completeness and accuracy of information processing is improved, but the prompt size exceeds token limits
Solution Approach 1:
The taxonomy is divided into multiple hierarchical levels (top-layer, intermediate layers, bottom-layer). Each level is processed separately in sequential prompts, allowing the complete taxonomy information to be incorporated without exceeding token limits in any single prompt.
Solution Approach 2:
The single-dimensional approach of including all taxonomy items in one prompt is transformed into a multi-dimensional hierarchical structure. The taxonomy is organized across multiple levels and processed through sequential prompts, effectively adding a temporal and structural dimension to the information delivery.
2Manufacturing precision
If hierarchical taxonomy structures are maintained, then the organizational accuracy and contextual relationships are improved, but the complexity of prompt construction and processing increases
Solution Approach 1:
The hierarchical taxonomy is segmented into distinct layers (top-layer, intermediate, bottom-layer), with each layer processed in separate prompts. This segmentation maintains the hierarchical structure and contextual relationships while simplifying the construction of individual prompts at each level.
Solution Approach 2:
The taxonomy is pre-processed and organized into hierarchical layers before being fed to the LLM. The top-layer taxonomy context is constructed in advance, and intermediate layers are iteratively processed, preparing the structured information beforehand to reduce processing complexity during execution.
3Productivity
If multiple taxonomy items are addressed in a single prompt, then the processing efficiency is improved, but the accuracy of matching specific elements decreases
Solution Approach 1:
The processing approach segments taxonomy items by hierarchical level rather than mixing all items in one prompt. Each layer is processed separately with focused attention, improving matching accuracy while maintaining overall efficiency through the structured sequential approach.
Solution Approach 2:
Instead of attempting to address all taxonomy items simultaneously, the system processes partial sets of items at each hierarchical level. This partial action approach ensures accurate matching for each subset while the cumulative effect across all layers achieves comprehensive processing.
Data Source
AI summary
Embodiments are directed towards a computer-implemented method for identifying one or more matching elements from a taxonomy present in a document. The method may include identifying the document, the taxonomy, and a question that may be processed by a large language model (LLM). The method may further include applying at least one taxonomy augmented generation (TAG) tactic from a set of TAG tactics to the identified document and to the identified taxonomy. The method may also include generating an input prompt that may be configured to be used as an input for the LLM, where the input prompt may include a document context derived from the document, a taxonomy context derived from the taxonomy, and the question to be processed by the LLM. The method may further include providing the input prompt to the LLM, and receiving a response generated by the LLM.