Emitted after a successful RAG retrieval during a turn so downstream
consumers (guardrails, grounding verification, UI source panels) can
see the retrieved chunks before the LLM call produces text deltas.
The chunk content is { ragSources: RagRetrievedChunk[] }.
One per model step that completes: the step's text, finish reason, provider, model, hop and usage.
A step that fails emits none: a USAGE_UPDATE with metadata.attemptFailed: true reports what it was
billed, when the provider reported that, and the turn's ERROR chunk follows. Content: StepFinishedChunkPayload.
One per result the GMI records for a call of its own tool round, failures included. Results a host
passes to handleToolResults() produce none. Content: ToolResultChunkPayload.
Defines the type of content in a
GMIOutputChunk.