Stardog Voicebox Release Notes

Release - Oct 17, 2024

• Updated Voicebox UI
  • The initial release of an updated Voicebox UI! The full-page Voicebox experience and chat windows from cloud.stardog.com, Explorer, and Designer are updated. Highlights include:
    • Modern splash page with data summary and spotlight questions
    • A human-user to computer message presentation rather than a conversation
    • New action icons, including bringing back both positive and negative feedback
    • Ability to hide “Explain response”

• [Private Preview] Enable additional Voicebox agents for additional capabilities. New agents are are as follows:
  • Analytics agent - Analyze the previously returned tabular result to answer questions about trends, comparisons, outliers using LLM’s inferencing capabilities.
  • Computation agent - Use LLM’s code writing capabilities to do numeric computations over the previously returned tabular result
  • Chart agent - Create a chart using the previously returned tabular result based on user’s instruction
  • External LLM agent - Use the LLM’s background knowledge to answer the question instead of querying the Knowledge Graph

  Note: At this time, Voicebox will not automatically use these agents unless the user provides a hint (see below) to use an agent specifically.

• Add support for hints to guide Voicebox when it makes mistakes. Hints are added to user questions with hashtags. For example:
  “#llm What is the capital of France?”
  Supported hints are grouped in two categories:
  • Hints to help Voicebox select the right agent
    • #analyze - Analyze the previously returned query results to answer the question
    • #chart - Create a chart using the previously returned query results
    • #compute - Perform a computation over the previously returned query results to answer the question
    • #llm - Use the external LLM to answer the question
    • #query - Query the knowledge graph to answer the question
    • #summary - Summarize the information about the entity mentioned in the question
  • Other hints
    • #nohistory - Ignore the conversation history while answering the question

  Note: The hints #analyze, #chart, #compute, #llm only works for users who are in the private preview for the new agents.

• Add more context to the response if Voicebox cannot answer a question
• Improve response times 20% by reducing agent selection and result summarization times
• Infrastructure changes to collect more detailed metrics and to allow faster diagnosis of issues
• Use asynchronous processes for Voicebox service to improve scalability
• Add support to use LLMs hosted by Databricks if configured

---