Profile-Memory Question Answering for Personalized LLM Responses
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models have limited memory capacity and provide one-size-fits-all answers, affecting user experience due to the inability to remember historical conversation messages and lack personalized reasoning capabilities.
Innovation Solution
A memory module is configured to store user profile messages and historical conversation messages, with lifecycle duration conditions and similarity thresholds, allowing the large language model to determine answer messages based on profile messages that match user queries, reducing memory usage by storing concise profile messages and extending relevant historical conversation messages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If an external memory module stores all historical conversation messages to improve memory capacity, then the large language model can remember more user interactions and provide personalized answers, but the memory space occupied by the memory module increases significantly
Solution Approach 1:
The patent extracts only the essential user profile information from historical conversation messages and stores it separately as profile messages, while the remaining detailed historical conversations are discarded. This extraction principle reduces memory space by keeping only the most important information (user profiles) that contributes to personalized reasoning capabilities.
Solution Approach 2:
The patent segments the memory storage into two distinct parts: profile messages stored in the memory module and detailed historical conversation messages not stored. This segmentation allows the system to maintain personalized reasoning capabilities through profile messages while avoiding the memory space consumption of storing all historical conversations.
2Adaptability or versatility
If the large language model stores and processes more historical conversation messages to improve personalized reasoning capabilities, then user experience improves with more accurate answers, but the input sequence length limit is exceeded
Solution Approach 1:
The patent extracts essential user characteristics from historical conversations and condenses them into concise profile messages. These profile messages capture the user's preferences, behaviors, and characteristics without including the full historical conversation content, thereby maintaining personalized reasoning while staying within input sequence length limits.
Solution Approach 2:
The patent transforms the representation of user information from detailed historical conversation messages to condensed profile messages with key attributes. This parameter change from storing complete conversation histories to storing essential user profile characteristics reduces the input sequence length while preserving the ability to provide personalized answers.
3Measurement precision
If the large language model is provided with all historical conversation messages to improve answer accuracy, then personalized reasoning is enhanced, but the model can only remember messages within the current session and provides one-size-fits-all answers
Solution Approach 1:
The patent performs preliminary processing of historical conversation messages to extract and store user profile information before the actual question answering process. By pre-extracting and storing essential user characteristics in profile messages, the system prepares personalized information in advance, enabling the model to provide accurate and personalized answers without being constrained by session memory limitations.
Data Source
AI summary
A question answering method, an electronic device, and a storage medium are provided in the present disclosure. The question answering method includes, based on a target query message inputted by a user, determining a first profile message in at least one profile message of the user stored in a memory module, where a similarity between the first profile message and the target query message is higher than a first specific threshold; and based on the target query message and the first profile message, using the large language model to determine an answer message of the target query message. At least a part of the at least one profile message stored in the memory module is obtained based on at least one historical conversation message, which satisfies a storage lifecycle-duration condition, of the user interacting with the large language model.