AQA RAG model JSON output and too high answer probability

Gemini API
1

AQA RAG Model Issues

When I try to use AQA model as outlined in the API docs, I couldn’t get a good response because 9/10 it wouldn’t adhere to the JSON output structure I outlined in the prompt. It doesn’t have a option for mime type: application/json so prompting was the best way. I am not sure if I am just missing some customization options for the AQA model or I am doing something wrong. Still, the AQA model doesn’t feel as though I have as much control over it in the sense of, what model is this using (Gemini 1.5 or 1.0) or is it its own standalone model? How do we customize and improve upon its responses? What is the cost structure for the AQA model?

Here is my code:

content = glm.Content(parts=[glm.Part(text=user_query)])
    retriever_config = glm.SemanticRetrieverConfig(source=corpus_resource_name, query=content)
    generate_answer_request = glm.GenerateAnswerRequest(
        model="models/aqa",
        contents=[content],
        semantic_retriever=retriever_config,
        answer_style=answer_style,
        temperature=0.2
    )
    aqa_response = generative_service_client.generate_answer(generate_answer_request)

Another issue I ran into was that the answerable_proability seemed a little too overconfident, and considering I’m using that value to determine whether to display the AQA model response or use general Gemini model, the value is kind of important. Is there any more details regarding answerable probability?