Chat Completions Functions - forcing it to get the parameters from the user?

I’m playing with the new function calling and its theoretically exactly what I need.

eg. I’m writing a murder mystery game and the player should be able to ask his assistant to obtain and execute search warrants. Before using functions, often (but not always) he would ask for probable cause or what the player was looking for in trying to execute the warrant.
That seemed like a nice idea so I wanted to keep that.
The assistant has a system prompts that contains the background that its a game and that he’s playing this character etc and he has a list of characters and locations that players might want him to do things to.
This is the functions declaration
[
{
“name”:“search_warrant”,
“description”:“obtain and execute a search warrant”,
“parameters”:{
“type”:“object”,
“properties”:{
“location_id”:{
“type”:“string”,
“description”:“the id of the location to search”
},
“probable_cause”:{
“type”:“string”,
“description”:“user provided probable cause required to obtain the warrant”
}
},
“required”:[
“location_id”,
“probable_cause”
]
}
}
]

As in the weather example, I expected it to ask a follow up question to get the probable cause, just like he did before. But instead, he’s making up the probable cause himself based on info in the system prompt (or possibly the messages I’ve coded in as his opening spiel to the character)

{
“id”:“chatcmpl-7RM4URSgNP7zbqtd1rNaAvZFpoOpL”,
“object”:“chat.completion”,
“created”:1686754054,
“model”:“gpt-3.5-turbo-0613”,
“choices”:[
{
“index”:0,
“message”:{
“role”:“assistant”,
“content”:null,
“function_call”:{
“name”:“search_warrant”,
“arguments”:“{\n "location_id": "9",\n "probable_cause": "To search for any evidence related to the murder of Mr. John Mulch."\n}”
}
},
“finish_reason”:“function_call”
}
],
“usage”:{
“prompt_tokens”:988,
“completion_tokens”:41,
“total_tokens”:1029
}
}

How can I make sure it asks the user for the probable cause or at least infers it from user messages provided in the context rather than from the system prompt or from messages in the context that come from the assistant?

Eg.