Live session systems manage active interactions between users and agents within context graphs. These systems maintain state awareness and handle real-time decision-making while users are actively engaged. The architecture supports various concurrent processes including policy enforcement, context management, and dynamic data integration.
Policy Management
During live sessions, the system continuously monitors for policy violations on a per-message basis. This real-time policy enforcement ensures compliance and safety throughout the interaction. The system must balance thorough policy checking with maintaining conversational flow, ensuring that safety mechanisms don't significantly impact the user experience.
Dynamic Context Management
Live sessions often involve the creation and management of child context graphs. For example, in mental health first aid scenarios, specialized context graphs may be instantiated to handle specific situations or interventions. These child graphs inherit relevant context from their parent while maintaining their own specialized state and objectives.
Real-Time Data Integration
The system supports dynamic tool calling for real-time data infusion during live sessions. This capability allows agents to access current information, perform calculations, or retrieve relevant data as needed during the interaction. The integration must be seamless and contextually appropriate, maintaining the flow of the session while enhancing its effectiveness.
Processing Characteristics
While live session systems operate in real-time, the processing speed varies based on the task complexity and requirements. Users understand they are in an interactive work session, and response times may range from seconds to minutes depending on the task. The system manages user expectations through appropriate feedback and progress indicators.
Post-Processing Systems
Independent Architecture
Post-processing systems operate independently from the agent and context graph architecture. These systems analyze completed sessions and generate various artifacts that inform future interactions. The separation from the live session architecture allows for more comprehensive and resource-intensive analysis without impacting user experience.
Memory Generation
After sessions conclude, post-processing systems analyze interaction patterns and outcomes to generate memory artifacts. These memories capture important insights, patterns, and learnings from the session that may be valuable for future interactions. The generation process can involve deep analysis of conversation flows, decision points, and outcomes.
User Model Updates
Post-processing systems update user models based on completed session data. This involves analyzing interaction patterns, preferences, and outcomes to refine the system's understanding of individual users. These updates help personalize future interactions and improve service delivery.
Task Analysis and Generation
The system analyzes completed tasks to identify patterns, bottlenecks, and opportunities for improvement. This analysis may lead to the generation of new task templates or the refinement of existing ones. The process involves understanding task structure, success patterns, and failure modes.
Metric Extraction
Custom metrics are extracted through post-session analysis, providing insights into system performance, user engagement, and outcome effectiveness. These metrics inform system improvements and help track progress toward business objectives. The extraction process can involve complex analysis of multiple session components and their relationships.
Processing Characteristics
Post-processing systems can perform deeper, more time-intensive analysis since they operate after sessions conclude. This allows for more thorough examination of session data and generation of high-quality artifacts. The processing time can vary significantly based on the analysis depth and complexity required.
System Interaction
Data Flow & Separation of Concerns
Data Flow
Live session systems produce raw interaction data, which post-processing systems analyze to create actionable insights and artifacts. These outputs inform future live sessions, creating a cycle of continuous improvement.
Separation of Concerns
Clearly separating live and post-processing systems ensures real-time responsiveness while enabling deep, resource-intensive offline analyses, maintaining optimal overall system performance.
Amigo’s distinct yet integrated system architectures—live sessions and post-processing—effectively combine real-time responsiveness with thorough analytical depth, optimizing both immediate interaction quality and long-term strategic improvements.
