idea

what

A knowledge graph containing a person’s

  • characteristics
  • life events
  • experience
  • traumas
  • thoughts
  • perspectives
  • traits
  • etc. and these things related to each other (e.g. how an event influence a personality trait).

application

Can be used to build virtual characters

  • Store the embeddings of each node in the graph.
  • On each query, retrieve the top k the closest nodes.
  • Extract the sub-graphs surrounding each retrieved node within a certain radius.
  • Process the extracted sub-graph to build context for generation.
  • Maybe fine-tune the LLM to generate using sub-graph results directly
    • process sub-graph to context as above
    • use the processed context for generation
    • use the generation output as training data