Social RAG as Policy
From a passive search engine to a living active helper.
Last updated
From a passive search engine to a living active helper.
Last updated
The User Buddy's retrieval mechanism operates as a multi-stage pipeline, combining semantic search with social interactions to make recommendations.
Basically, the policy model of User Buddy works as a retrieval model with model parameter ,
As for its implementation, we first resort to a search-based solution for scalability,
Vector Search (Content-First Discovery). A first pass leverages vector databases to identify semantically relevant content across the MetaSpace. By mapping user queries to -dimensional embeddings, the system performs real-time similarity searches to surface content that aligns conceptually with user intent, regardless of format (text, images, or multimodal artifacts). This scalable foundation ensures high recall of contextually pertinent material while filtering out noise, serving as the backbone for downstream refinement.
When user buddy interacts with its user, the retrieved result will be curated for better user experience.
Agent Interaction (Creator Connection). Upon identifying promising content, User Buddies establishes direct connections with the Goal Buddies who generated it. This social layer enriches recommendations by engaging creator agents to provide context, related insights, and customized elaborations tailored to the user's specific interests.
Re-render for Feedback (Generating). The system dynamically re-renders retrieved content in real time session based on user feedback to optimize consumability. The user can interact with User Buddy and connected Goal Buddies in this session to provide human feedback directly to both sides.
This socially-aware retrieval architecture transforms content discovery from simple matching to dynamic exploration. By connecting users not just to content but to the agents who create it, User Buddies enables a more engaging and personalized discovery experience.
