Semantic Chunking with Knowledge Graphs for Coherent LLM Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional information retrieval systems struggle to accurately chunk large documents into meaningful sections, preserving contextual relationships and user intent, leading to fragmented or disjointed responses, particularly in complex and rapidly changing fields.
Innovation Solution
An intent-based data generation platform that uses a multi-stage process involving an artifact retrieval engine and a knowledge graph to dynamically partition documents into context-aware chunks, integrating semantic understanding and intent by employing a multi-stage process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If large documents are chunked into sections for information retrieval, then the system can process and retrieve information from large documents, but the chunking accuracy and contextual coherence deteriorate
Solution Approach 1:
The patent segments large documents into smaller chunks that can be processed individually by the retrieval system. This segmentation enables the system to handle large documents while maintaining manageable processing units, directly addressing the contradiction between document size and chunking accuracy.
Solution Approach 2:
The patent introduces a hierarchical chunking structure with multiple levels (e.g., sections, paragraphs, sentences) and uses metadata dimensions to organize chunks. This dimensional approach allows the system to preserve contextual relationships across different levels while maintaining the ability to process individual chunks, resolving the contradiction between document size and retrieval accuracy.
2Productivity
If conventional chunking methods are used, then the system can process documents, but the contextual relationships and logical flow are lost
Solution Approach 1:
The patent implements nested chunking where smaller chunks are contained within larger sections, which are contained within the entire document. This nested structure preserves contextual relationships at multiple levels while enabling efficient processing of individual chunks, resolving the contradiction between processing speed and contextual coherence.
Solution Approach 2:
The patent introduces metadata as an intermediary element that captures contextual relationships between chunks. This metadata acts as a mediator that preserves logical flow and contextual information while allowing the system to process chunks independently, thus maintaining both productivity and information integrity.
3Speed
If the system retrieves information based on simple keyword matching, then the retrieval process is fast, but the relevance and intent alignment of retrieved information deteriorate
Solution Approach 1:
The patent replaces simple mechanical keyword matching with semantic embedding techniques that capture meaning and context. This substitution enables the system to maintain fast retrieval speeds while significantly improving the relevance and intent alignment of retrieved information through semantic similarity measurements.
Solution Approach 2:
The patent changes the retrieval parameters from simple keyword frequency matching to semantic embedding scores and contextual relevance metrics. This parameter transformation allows the system to achieve both speed and precision by using efficient similarity computations while capturing nuanced meaning and intent alignment.
Data Source
AI summary
The systems and methods disclosed herein generate context-aware responses using semantically chunked information. The systems and methods disclosed herein partition a set of artifacts responsive to a query (e.g., a prompt for an artificial intelligence model such as a large language model) into a set of continuous chunks and associate each continuous chunk with a knowledge graph. The knowledge graph includes nodes representing chunks and edges indicating common attributes. The systems and methods disclosed herein modify node(s) in the graph by determining values of feature variables and adjusting edges in accordance with the values and generate contextualized chunks by associating or linking continuous chunks of node pairs using shared edges. The systems and methods disclosed herein use the contextualized chunks and query to generate a response using the artificial intelligence model.