We want to be able to pass a param saying to the agent that he cannot execute a set of actions or actions including a set of keywords. If those base actions are part of the BaseAction they are simply filtered out of the input sent to the LLM (the LLM is blind of that action) and if it's a more complicated set of actions eg. "buttons with text delete" then this needs to be handled either A) with a fine-tuned prompt or B) at LLM output, and if LLM selected an unauthorized action a guided retry system
We want to be able to pass a param saying to the agent that he cannot execute a set of actions or actions including a set of keywords. If those base actions are part of the
BaseActionthey are simply filtered out of the input sent to the LLM (the LLM is blind of that action) and if it's a more complicated set of actions eg. "buttons with text delete" then this needs to be handled either A) with a fine-tuned prompt or B) at LLM output, and if LLM selected an unauthorized action a guided retry system