Binary Hypervector Retrieval Graphs for Scalable RAG Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional retrieval-augmented generation systems face high computational complexity and memory usage issues due to real-valued high-dimensional embedding spaces, limiting scalability and response time in large-scale document corpora.
Innovation Solution
A hyperdimensional computing-based framework converts transformer-based token embeddings into binary hypervectors, constructing a target graph in a retrievable structure to maintain document relationships, reducing memory usage and improving scalability and response speed through integer bit-level operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real-valued high-dimensional embedding space is used for document representation, then retrieval accuracy can be maintained, but memory usage and computational complexity increase significantly
Solution Approach 1:
The patent changes the parameter type of embedding vectors from real-valued to binary hypervectors. This transformation maintains the ability to represent document semantics while dramatically reducing memory requirements, as binary vectors require far less storage space than their real-valued counterparts.
Solution Approach 2:
The patent creates binary hypervector representations as simplified copies of real-valued embeddings. These binary copies preserve the essential semantic information needed for retrieval while using significantly less memory, allowing the system to work with compressed versions of the original high-dimensional data.
2Loss of information
If real-valued high-dimensional embedding space is used for document representation, then semantic relationships can be captured, but computational speed decreases due to complex operations
Solution Approach 1:
The patent replaces complex real-valued mathematical operations with simpler binary operations. By substituting continuous arithmetic computations with discrete binary hypervector operations, the system achieves faster processing speeds while maintaining semantic relationship preservation through the structured design of binary hypervector algebra.
3Reliability
If brute-force search is performed across all documents in large-scale corpus, then comprehensive search coverage is achieved, but response time becomes unrealistic
Solution Approach 1:
The patent segments the large-scale document corpus into manageable clusters represented by binary hypervector centroids. Instead of searching through every individual document, the system divides the search space into regions, each represented by a centroid, allowing efficient navigation and retrieval without sacrificing comprehensive search coverage.
Solution Approach 2:
The patent performs preliminary organization of documents into clustered structures with binary hypervector representations before retrieval operations. This pre-processing creates an efficient index structure that enables fast search operations, avoiding the need to scan all documents during actual retrieval tasks.
4Productivity
If vector indexes are built for large-scale document retrieval, then search efficiency is improved, but memory requirements increase significantly
Solution Approach 1:
The patent changes the parameter representation in vector indexes from real-valued vectors to binary hypervectors. This parameter transformation maintains index functionality for efficient search operations while dramatically reducing the memory footprint of the index structure, enabling scalable deployment on systems with limited memory resources.
Data Source
AI summary
Disclosed is a hyperdimensional computing-based retrieval-augmented generation apparatus. The retrieval-augmented generation apparatus may include a network interface; at least one processor operatively connected to the network interface; and at least one memory operatively connected to the at least one processor, and the at least one memory may store instructions that, when executed, cause the at least one processor to generate a target binary hypervector through hyperdimensional computing (HDC)-related operation for a target document, to generate a target graph by connecting at least one binary hypervector included in a hyperdimensional space in a retrievable structure based on identifying the hyperdimensional space in which the target binary hypervector is included, and to determine a retrieved document similar to a query hypervector in the hyperdimensional space by applying the query hypervector to the target graph based on receiving the query hypervector.