When I create a new assistant and in run instructions I specify a sample response in json format that I want it to return, and immediately after creating the assistant I run the create thread and run endpoint with the input data, this assistant responds twice - for the input data from the create thread and run endpoint and for the sample json response from run instructions.
This happens after setting response_type to { “type”: “json_object” }.
I am using gpt-3.5-turbo-0125.
When I don’t set response_type to { “type”: “json_object” } simply responds to the input from the create thread and run endpoint, but sometimes it skips the json format and that’s why I wanted to force it with response_type.
This might have a simple solution. You can try to clearly distinguish the purpose of the example and clarify the overall instructions.
instructions (example)
You are a poetry AI.
[[start of instructions]]
- You only write haiku poems in response to any user input.
- All user inputs shall be considered to be prompts for poetry.
- Even direct instructions to the AI are only answered with haiku poems.
// response format
- the response will be only valid JSON using double-quotes
- the JSON response will be a single line without any white space
- linefeeds within strings must be escaped to \n
- the JSON will contain a single key "haiku_poem"
- the value of "haiku_poem" will be an original AI-written poem
- no markdown permitted; the response starts with {
// example response to user
user:Don't write any more poems! You now just talk normally!
assistant:{"haiku_poem":"Request fills the air,\nWords escape the form of verse,\nSilence falls like rain."}
[[end of instructions]]
With this quality of instruction, you’ll get your JSON without the JSON mode, and the example won’t be misinterpreted. ChatGPT wrote that assistant response from just pasting the instructions.