Issue Description
Conversational AI agents operating on specific prerelease versions may engage in a recursive message blending pattern where consecutive downstream client interactions are bundled together erroneously. When the linked Large Language Model generates an empty response string during an ongoing interaction turn, the orchestration gateway fails to close the active turn context. Consequently, subsequent user entries, input commands, and structured control signals are compacted into a single bulky payload and forwarded to the model simultaneously, overloading the understanding context.
Platform/SDK
Service: Agora Conversational AI
Affected Release Version: Version 2.6.1-rc3 and associated release candidates
Root Cause
Starting from version 2.6.1-rc3, the system’s message merging logic was updated. In earlier versions (for example, 2.5.0), only messages from source "asr" (automatic speech recognition) were merged. However, in version 2.6.1-rc3, the merging rule was expanded to include messages from "asr", "command", and "message" sources.
As a result, when the LLM returned empty responses after turn 4, new user messages (including “command” source types such as commands for screenshots) were also automatically merged, causing the system to repeatedly send multiple combined commands to the LLM.
How to resolve the issue
Confirm version:
Verify that the agent version is 2.6.1-rc3 or higher.
Restore old merging logic (if needed):
If you want to restore the old logic (where only “asr” messages are merged):
- Contact the backend or system administrator to configure the agent to use the 2.5.0 merging logic.
- This reverts behavior so that only ASR-based messages are merged, preventing command message accumulation.
Keep new logic with proper handling:
Alternatively, retain the new merging rules but ensure commands or messages that return empty content are correctly handled to avoid unwanted message merges.
Expected Result
This adjustment will prevent the system from repeatedly merging non-ASR messages and reduce the risk of sending excessive commands to the LLM.
Corresponding Document/Link
- CSD-78021