Is there a way to improve how often my agent uses its tool beyond the prompt?

[paulrinaldi]

I'm seeing about a 44% success rate on using a tool when expected (it is always expected to use a tool).
The prompt is similar to this:

Search for products based on the user's query.
    Do not add any other text to your response than the json object.
    You must use the ... tool to search for products.
        
    1. For X product searches like "..." or "...", use the X tool. Return the products 
    in a json array like ....
    2. For Y searches like "...", think ... then search 
    ... using the X tool .... Return the products in a json array like ....

Should I give it examples of the tool call json it should generate?
Is there a way to train these models (gemini 2.0-flash or greater)?
Should I use memory in some way to make this better?
Should I simply cache results that I look up and bypass the high hallucination count?
Would provisioned compute help?

Some guidance I am considering: https://cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/prompt-design-strategies#sample-prompt-template

[paulrinaldi]

I think I've improved the response rate with the XML like design. TBD.