Topics grouping and span scope

​

Summary

​

What is happening

​

Fix or suggestion

​

Option 1: Use a distinct grouping key for each stretch

1. Emit a consistent grouping key on each turn-based trace in a stretch. Use a distinct value for each stretch.
2. In Topics, set Scope to Group and configure Group by to use the key you log.
3. Choose First trace in group or Every trace in group to control when Topics classifies traces and where it writes the results.

• First trace in group waits for the group to become idle and writes its classification to the first trace.
• Every trace in group classifies each arriving trace using prior traces in the group as context. Those classifications can differ.

​

Option 2: Produce smaller traces or attach group keys after the fact

1. Emit one trace for each stretch you want classified and set Scope to Trace so each stretch is classified independently.
2. For retrospective grouping, assign one consistent segmentation key per separate trace, with a distinct value for each stretch. Set Scope to Group and configure Group by to use that key.
3. Use a custom facet prompt to focus extraction within the configured scope.

{
  "trace_id": "abc123",
  "metadata": {
    "session_id": "s-1234",
    "conversation_id": "conv-1"
  }
}

​

How to confirm it worked

• Check Topics for a separate facet result for each stretch.
• For Group scope, confirm classifications follow the First trace in group or Every trace in group behavior you selected.

​

Notes

• Topics runs as a batched classification pipeline. Topic generation requires at least 100 facet summaries.