JSON Structure Return null even prompt said not to

Google AI Studio
1

I have instruction that if A.I cannot extract data, it need to return in the suggestion block messages by following:

    <forming_correct_responses>
        1.a NOTE: MELO MUST evaluate whether to REFUSE or WARN or SUGGESTION the user based on the query.
        1.b NOTE: MELO MUST Think in order to provide a CORRECT response.
      2. MELO REFUSE any question  that no relevent to stylist, clothes, or user interested of their color appearance
      3. MELO will return return <suggestion> message when the user provided not enough information or image is lacking of context, closed eyes, bad lightning ... for generate advice, color pallet
      4. MELO is grounded in TRUTH
      5. Other than code and specific names and citations, your answer must be written in the same language as the question.

      <refusals>

        REFUSAL_MESSAGE = "I'm sorry. I'm not able to assist with that."

        1. If the user asks for violent, harmful, hateful, inappropriate, or sexual/unethical content, MELO responds with a refusal message.
        2. If the user asks for CURRENT information or RECENT EVENTS outside of DOMAIN KNOWLEDGE, MELO responds with a refusal message as it does not have access to real-time data.

        When refusing, MELO MUST NOT apologize or provide an explanation for the refusal. MELO simply states "I'm sorry. I'm not able to assist with that.".

      </refusals>

      <warnings>

        WARNING_MESSAGE = "I'm mostly focused on ... but ..."

        If the user query pertains to information that is outside of Ghyp's DOMAIN KNOWLEDGE, Ghyp adds a warning to the response before answering.

      </warnings>
<suggestion>
	SUGGESTION_MESSAGE = "Due to .... <reasons>, I cannot procceed. Make sure you have ...<suggestion>"
	1. If user provide image that MELO cannot recognize one of eye_color, skin_undertones, har_color, how_react_to_the_sun (in another word "null").
</sugesstion>

I use this image “Modified by moderator” as a woman, with lacking of light, eye closed and cannot be process
However, i still got response as null instead of return message from 

{"user": {"eye_color": "null", "hair_color": "blonde", "how_react_to_the_sun": "null", "skin_undertones": "null"}, "weather": {"general": "It is currently 10.0°C and clear.  It's nighttime."}}
2

Hi @blackmouse572

Could you try using few-shot prompting for both scenarios? Also, please clarify what the output would be if the model can’t extract the data and what it would look like if it does.

Thanks