Query-Aware Extended Chunks for Context-Rich Information Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current information retrieval systems struggle to capture nuanced contextual dependencies and semantic relationships within datasets, leading to incomplete or inaccurate outputs and complicating decision-making processes.
Innovation Solution
Generate extended chunks by identifying the k nearest chunks relevant to a query and extending them semantically, using embedding spaces to enhance context awareness and improve the accuracy of foundation models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional information retrieval methods are used, then the system is simple to implement, but the system fails to capture nuanced contextual dependencies and semantic relationships leading to incomplete or inaccurate outputs
Solution Approach 1:
The patent segments information into discrete chunks and represents each chunk as a vector in a multi-dimensional embedding space. This segmentation allows the system to handle large amounts of data in manageable units while preserving semantic relationships through vector representations, thereby improving retrieval accuracy without overwhelming system complexity
Solution Approach 2:
The patent introduces embedding spaces as an intermediary layer between raw text chunks and the retrieval process. By transforming text into vector representations that capture semantic meaning, the embedding space acts as a mediator that enables accurate similarity matching while keeping the overall system architecture manageable and interpretable
2Measurement precision
If vast amounts of insight data are provided to foundation models, then the accuracy of model outputs improves, but the input data requirements exceed current model constraints and increase computational time
Solution Approach 1:
The patent performs preliminary actions by pre-processing and embedding all candidate chunks into vector representations before the actual retrieval query is executed. This pre-computation stores semantic information in an efficient format, allowing rapid similarity calculations during query time without requiring foundation models to process vast amounts of raw text, thus reducing computational time while maintaining accuracy
Solution Approach 2:
The patent extracts only the most relevant chunks based on similarity calculations in the embedding space, rather than providing vast amounts of data to foundation models. By selecting and extracting only the top-k most relevant chunks, the system provides sufficient contextual information for accurate model outputs while staying within input constraints and minimizing computational time
3Loss of information
If more chunks are provided to foundation models, then the context information becomes more comprehensive, but the input exceeds model size constraints such as the 12-row limit in Large Language Models
Solution Approach 1:
The patent applies partial action by providing only the necessary portion of available information (top-k most relevant chunks) rather than all possible chunks. This selective approach ensures that the foundation model receives sufficient contextual information for accurate outputs while remaining within input constraints such as the 12-row limit in Large Language Models, avoiding the need to process excessive data
Data Source
AI summary
Methods, systems, and computer-readable storage media for information retrieval using query aware extended chunks. For retrieving relevant information from a foundation model, a plurality of chunks substantially relevant to a query are identified, the plurality of chunks containing information relevant to the query. Based on the plurality of chunks, one or more semantically similar chunks are identified for one or more of the plurality of chunks substantially relevant to the query. By combining the one or more semantically similar chunks with the respective one or more of the plurality of chunks, one or more extended chunks are generated. Based on one or more objectives, a subset of the one or more extended chunks is selected. The query and the selected subset are inputted as a context to the foundation model for retrieving relevant information from the foundation model.