`corrected_input` by SelectorAgent includes system messages and may be truncated

[mingxin-zheng]

Description

The output of SelectorAgent, corrected_input, is supposed to be a cleaned version of ASR text to helps subsequent agents in the workflow to process user's requests

However, the actual output can contain

• Body text from system message: You are an intelligent laparoscopic cholecystectomy surgical assistant who must select from the available agents based on the user's request.
• Truncated output if the user message is long, e.g. ending a sentence with "..."

Reproduce

With the uncertainty of LLMs in mind, one can try prompts like "Check the text in the image" / "Repeat the text". The VLM may be triggered to provide incorrect corrected_output

Alternative Solution

Engineered prompts of the subsequent agents, and let those agents know the user text can be from ASR and may contain errors.