Generative Graph Retrieval for Connecting Dispersed Information
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional information retrieval systems struggle with cluttered datasets where relevant information is dispersed across different documents or sections, lacking the ability to effectively identify and connect dispersed pieces of information, leading to incorrect zero results, especially in complex queries requiring reasoning and derivation.
Innovation Solution
A generative, graph-enhanced information retrieval approach that creates a structured graph representation of documents, using embeddings and graph-based retrieval to connect nodes and extract paths responsive to queries, incorporating a generative AI model to generate answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional information retrieval systems are used on cluttered datasets, then the system structure remains simple, but the ability to identify and connect dispersed information deteriorates, leading to incorrect zero results
Solution Approach 1:
The patent transforms flat document data into multi-dimensional graph structures with nodes representing entities, concepts, or documents and edges representing relationships. This dimensional transformation enables the system to capture dispersed information connections that conventional linear retrieval methods miss, thereby improving retrieval accuracy without proportionally increasing system complexity
Solution Approach 2:
The patent introduces graph embedding technologies as intermediaries that transform graph structures into vector representations. These embeddings serve as mediators between the complex graph data structure and the retrieval algorithm, enabling efficient similarity search and connection identification while maintaining system tractability
2Reliability
If graph-based retrieval methods are implemented to connect dispersed information, then the ability to retrieve complex information improves, but the computational complexity and processing time increase
Solution Approach 1:
The patent pre-computes graph embeddings for all nodes in the knowledge graph during an offline phase. This preliminary action transforms the complex graph structure into ready-to-use vector representations, enabling fast online retrieval without repeatedly computing graph relationships during queries, thus reducing processing time while maintaining completeness
Solution Approach 2:
The patent creates vector space copies of the graph structure through embedding techniques. Instead of querying the original complex graph directly, the system queries the embedded vector space which preserves structural relationships but enables efficient computational operations, effectively copying the graph's informational content into a more tractable format
3Measurement precision
If detailed graph structures are created to represent document relationships, then the precision of information connection improves, but the device complexity and storage requirements increase
Solution Approach 1:
The patent transforms discrete graph parameters (nodes, edges, relationships) into continuous vector parameters through embedding. This parameter transformation maintains the precision of information connections by preserving relational semantics in the vector space, while the continuous nature of vectors enables more efficient storage and computation compared to explicit graph structures
Data Source
AI summary
Embodiments of the disclosed technologies include parsing a query into a first query portion and at least one second query portion, matching an embedding of the at least one second query portion with an embedding that corresponds to a portion of a document of a document set, mapping the portion of the document to a first node of a graph; by a generative artificial intelligence model, constructing a graph query based on at least the first node, executing the graph query on the graph to identify a second node of the graph, extracting a path from the graph, and configuring the path for output at a device.