RAG Knowledge Acquisition for Edge AI Classification Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Discriminative AI systems struggle when training datasets cannot describe the complete scenario, while generative AI systems face issues like hallucinations and high parameter requirements, limiting their effectiveness in complex scenarios.
Innovation Solution
A hybrid distributed inference system combining discriminative and generative AI models, with localized small foundation systems and centralized large foundation systems, utilizing retrieval augmented generation (RAG) to optimize classification and reduce computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If generative AI systems are used to generalize to data not in training set, then adaptability is improved, but hallucinations and large number of parameters increase system complexity
Solution Approach 1:
The system segments the AI architecture into multiple specialized components: discriminative models for classification, generative models for data synthesis, and RAG modules for knowledge retrieval. Each component has a specific function and manageable parameter count, avoiding the need for a single large generative model while maintaining generalization capabilities.
Solution Approach 2:
The patent introduces retrieval-augmented generation (RAG) as an intermediary mechanism between the model and external knowledge sources. This allows the system to generalize to unseen data by retrieving relevant information from training data rather than relying solely on internal parameters, reducing the parameter burden.
2Manufacturing precision
If discriminative AI systems are used for classification, then manufacturing precision is improved, but reliability deteriorates when training dataset cannot describe complete scenario
Solution Approach 1:
The system merges discriminative and generative AI models into a hybrid architecture. The discriminative model provides accurate classification for known scenarios, while the generative model synthesizes relevant training examples to improve generalization to unseen scenarios, combining the strengths of both approaches.
Solution Approach 2:
The system performs preliminary actions by generating synthetic training data and retrieving relevant examples before classification. This prepares the system to handle unseen scenarios by pre-synthesizing relevant training cases and making them available when needed, improving reliability without compromising classification precision.
3Adaptability or versatility
If centralized large foundation systems are used, then adaptability is improved, but data transmission requirements and computational resources increase
Solution Approach 1:
The patent implements local small foundation systems at edge sites that are specialized for local conditions and requirements. These local systems handle most classification tasks independently, reducing the need for centralized processing and data transmission. Only when local systems encounter uncertain cases do they interact with the centralized system, minimizing data transmission volume while maintaining adaptability through localized intelligence.
4Productivity
If local small foundation systems are used at edge sites, then productivity is improved, but device complexity increases due to distributed architecture
Solution Approach 1:
The system implements a dynamic distributed architecture where local and centralized components interact based on real-time needs. The architecture adapts its complexity by engaging only the necessary components for each task - local systems handle routine classifications independently, while the centralized system provides support only when needed, optimizing productivity without requiring full distributed complexity to be active simultaneously.
Data Source
AI summary
The disclosure describes a retrieval augmented generation (RAG) knowledge acquirer. The RAG knowledge acquirer receives output data from a local small foundation system and a centralized foundation system both of which use a RAG system for a generative AI model. The RAG knowledge acquirer creates new enterprise data for an enhanced distributed RAG system used by the local small foundation system from the data. The new enterprise data is used to create query optimized representations (QORs). The RAG knowledge acquirer is connected to an external information repository allowing data sharing and federated learning with the other edge sites. The new enterprise data incorporates knowledge gained from previous inferences made by the local small foundation system and the centralized foundation system and improves the function of the local small foundation system keeping it customized and avoiding finetuning. The new enterprise data can also verify promoted local QORs.