AI Assistant Knowledge Set Selection for Low-Hallucination Responses
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current large language models (LLMs) for AI assistants are resource-intensive, requiring vast data and cloud-based processing, leading to inefficiencies and increased chances of hallucinations due to excessive information access.
Innovation Solution
Implementing small, offline LLMs with managed knowledge sets that are dynamically loaded based on user characteristics, instrument specifics, and regulatory compliance, using a master vector database to retrieve relevant information, thereby reducing unnecessary data access and minimizing hallucinations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current large language models are used for AI assistants, then comprehensive information access is achieved, but resource consumption and hallucination risk increase
Solution Approach 1:
The patent segments the knowledge base into multiple specialized vector databases organized by domain, topic, or data type. Instead of using a single large LLM with access to all information, the system divides knowledge into manageable chunks that can be selectively retrieved and processed, reducing the computational burden while maintaining comprehensive information coverage.
Solution Approach 2:
The patent extracts only the necessary knowledge subsets required for each specific query from the larger knowledge base. By using embedding models to identify and extract relevant vectors from specialized databases, the system retrieves only the essential information needed to answer a query, avoiding the resource-intensive processing of unrelated data while minimizing hallucination risks.
2Manufacturing precision
If vast quantities of data are provided to LLMs, then model training completeness improves, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary organization of knowledge data into specialized vector databases with proper embedding and indexing before the LLM training process. By pre-structuring the knowledge base into domain-specific collections with optimized embeddings, the system enables more efficient training convergence and faster inference, reducing both training time and computational resource requirements while maintaining model quality.
Data Source
AI summary
A test and measurement instrument includes one or more memories, a generative artificial intelligence (AI) model to access to the one or more memories, a display, user controls, and one or more processors configured to access an application programming interface (API) of an AI assistant for the generative AI model, receive one or more user inputs as a prompt through one or more of the API or a user interface, access a master vector database to retrieve a list of master candidates, compare the prompt to the list of master candidates to select ones of the master candidates, send the selected ones to a vector database, receive specific candidates from the vector database, send the prompt and the specific candidates to the generative AI model, receive a response, and display the response on the display.