Classifier-Routed Retrieval Augmented Generation for Multi-Domain Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Retrieval-Augmented Generation (RAG) processes struggle with complex data storage systems, including multiple data domains, formats, access permissions, and discriminating between prompts that require data retrieval versus LLM generation alone, leading to inefficient and inappropriate responses.
Innovation Solution
A method to generate a training dataset for a classifier that determines the appropriate retrieval process from a plurality of processes based on input prompts, using embeddings and annotations to distinguish between different data sources, formats, access permissions, and retrieval methods, and employs a trained classifier to select the optimal retrieval process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single data source is provided to RAG algorithm for organizational data, then comprehensive data access is enabled, but data retrieval accuracy deteriorates due to inability to handle multiple data domains and formats
Solution Approach 1:
The patent segments the monolithic data source into multiple data domains (e.g., structured data, unstructured data, semi-structured data) with distinct access methods. Each domain is treated as a separate retrieval target, allowing the system to selectively query only relevant domains based on the input prompt, thereby improving retrieval accuracy while maintaining comprehensive access capability.
Solution Approach 2:
The patent introduces an intermediary classification layer between the RAG algorithm and the data sources. This classifier determines which data domains are relevant to the given prompt and routes queries accordingly, enabling accurate retrieval from complex multi-domain data structures without requiring the RAG algorithm to directly handle all data types.
2Quantity of substance
If all data formats are made accessible to RAG process, then data completeness is improved, but retrieval complexity increases due to different formats requiring different retrieval methods
Solution Approach 1:
The patent creates a universal retrieval framework that can handle multiple data formats (structured, unstructured, semi-structured) through a common interface. The system uses format-detection capabilities and adaptive retrieval strategies that automatically adjust to the data format, allowing comprehensive data access without proportionally increasing retrieval complexity.
3Loss of information
If RAG process accesses entire data set, then information completeness is improved, but processing time increases due to unnecessary data retrieval
Solution Approach 1:
The patent implements partial retrieval by using the classifier to identify and query only the subset of data domains relevant to the specific prompt. Instead of retrieving from the entire data set, the system performs targeted queries on selected domains, reducing processing time while maintaining information completeness for the given context.
4Device complexity
If no classification is performed on input prompts, then system simplicity is maintained, but response appropriateness deteriorates due to inability to distinguish between prompts requiring retrieval versus LLM generation
Solution Approach 1:
The patent applies preliminary classification to input prompts before initiating the RAG process. The classifier analyzes the prompt characteristics and determines whether data retrieval is necessary and which data domains are relevant. This preliminary step ensures response appropriateness by preventing unnecessary retrieval operations and directing prompts to the most suitable processing path.
Data Source
AI summary
Methods and apparatus for generating a response to an input prompt are provided, in which a classifier is used to determine a retrieval process, from a plurality of retrieval processes, for use in generating a response to the input prompt. Methods and apparatus are also provided for training a classifier for determining a retrieval process, from a plurality of retrieval process and for generating a training dataset for training the classifier.