Collaborative LLM Session Merging for Context Accuracy and Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems utilizing large language models are inflexible, inefficient, and insecure, particularly when handling multi-user interactions and context recall, as they fail to adapt to multiple users, require excessive computational resources, and lack data privacy distinctions.
Innovation Solution
A collaborative interpreter system that merges sessions of a large language model by combining the states of an agent using a context engine, generating a merged prompt and state to facilitate multi-user interactions, reduce computational expense, and ensure data privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing systems tokenize the entire conversation history for context recall, then context accuracy is improved, but computational expense increases exponentially
Solution Approach 1:
The patent segments the conversation history into individual turns and selectively tokenizes only relevant portions for context recall, rather than processing the entire conversation history. This segmentation approach maintains context accuracy by preserving important information while dramatically reducing computational expense by avoiding redundant tokenization of all previous turns.
Solution Approach 2:
The patent extracts and stores context information from individual conversation turns in a structured format that can be directly reused. By taking out only the essential context elements needed for subsequent interactions, the system avoids the computational burden of re-tokenizing entire conversation histories while maintaining accurate context retrieval.
2Loss of information
If existing systems store all user account data indiscriminately, then data availability is improved, but security is worsened due to inability to distinguish private from shareable data
Solution Approach 1:
The patent applies local quality by assigning different access permissions and privacy levels to different portions of user account data. Instead of treating all data uniformly, the system tags and manages data with specific privacy attributes, allowing the large language model to access only the appropriate subset of data for each interaction, thereby maintaining both availability and security.
Solution Approach 2:
The patent introduces an intermediary layer (the system architecture with permission checking mechanisms) that sits between the stored user data and the large language model. This intermediary filters and manages data access based on privacy settings, ensuring that private data remains protected while shareable data is appropriately made available to the model and other users.
3Device complexity
If existing systems are designed for single-user interactions, then system simplicity is maintained, but adaptability to multi-user sessions is lost
Solution Approach 1:
The patent implements universality by designing the system to handle both single-user and multi-user interactions through a unified architecture. The system uses user identification and session management mechanisms that work seamlessly for individual users while also supporting collaborative multi-user sessions, allowing the same core system to adapt to different interaction modes without requiring separate specialized systems.
Solution Approach 2:
The patent applies dynamics by making the system's user session handling flexible and adaptive. The system can dynamically adjust between single-user and multi-user modes based on the interaction context, using dynamic user identification and session state management that automatically accommodates the appropriate number of users without requiring predetermined configuration changes.
Data Source
AI summary
The present disclosure relates to systems, non-transitory computer-readable media, and methods for combining sessions of a large language model. For instance, the disclosed systems can receive a first query from a first collaborating user account and a second query from a second collaborating user account. In one or more cases, the disclosed systems can generate a merged prompt based on the from the first query and second query. In some implementations, the disclosed systems can generate a response to the first query and/or the second query by processing the merged prompt utilizing the large language model.