Generative Model Whiteboard for Context Memory Fidelity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) face limitations in storing and retrieving user interaction history due to information loss and difficulty in locating relevant information using conventional vector search techniques, leading to potential loss of fidelity in synthetic memory extraction and retrieval.
Innovation Solution
Incorporating a whiteboard mechanism into generative models to store and manage user interaction history, allowing for continuous updating and archiving based on user interactions, and enabling user modification and display of relevant information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If the size of user interaction history passed to the LLM is increased to retain more information, then information completeness is improved, but the model's context window limit is exceeded causing information loss
Solution Approach 1:
The patent segments the user interaction history into two parts: a compressed summary stored in the whiteboard and the full detailed history stored externally. The whiteboard maintains a condensed version with key information while the complete interaction history is preserved separately, allowing the model to access both compressed and detailed information without exceeding context window limits.
Solution Approach 2:
The whiteboard acts as an intermediary between the user interaction history and the LLM. It receives the full interaction history, processes and compresses it into essential information, and presents this condensed version to the model. This intermediary layer prevents information loss by maintaining both compressed and detailed versions of the history.
2Productivity
If synthetic memory extraction and retrieval is used to manage interaction history, then information retrieval efficiency is improved, but fidelity is lost due to reformulation and consolidation
Solution Approach 1:
The whiteboard creates and maintains a copy of the user interaction history in a compressed format. This copy preserves the essential information and key details while being more compact. The system can retrieve information from this copy efficiently while maintaining fidelity to the original interaction history.
3Ease of operation
If conventional vector search techniques are used to locate relevant information, then search simplicity is maintained, but information location accuracy deteriorates
Solution Approach 1:
The whiteboard segments the interaction history into organized sections with clear structure and headings. This segmentation makes it easier to locate specific types of information by navigating to relevant sections rather than searching through unstructured text, improving both simplicity and accuracy of information location.
Data Source
AI summary
A computing system is provided, including processing circuitry configured to cause an interaction interface for a trained generative model to be presented, in which the interaction interface is configured to communicate a portion of a user interaction history. The processing circuitry is further configured to receive, via the interaction interface, an input for the trained generative model to generate an output. The processing circuitry is further configured to send a command to create, via the trained generative model or another trained generative model, a whiteboard based on the user interaction history and receive the created whiteboard. The processing circuitry is further configured to generate a prompt based on the whiteboard and the instruction from the user and provide the prompt to the trained generative model. The processing circuitry is further configured to receive a response from the trained generative model and output the response via the interaction interface.