Completions API response always considerably shorter than ChatGPT response

Prompting
1

I am having a hard time getting the completions API to return a detailed response. I have tried specifically asking for numbers of words, asking for 3-4 paragraphs per section. I have said all variations of requesting a verbose response with lots of details and be informative.

Every single time, it comes back with a response that is around 3000 characters. If i copy and paste the exact same prompt into chatGPT (chat dot openai dot com), it will generate a response that is typically twice the size of the response i got back from the completions API.

I have tried using gpt3-turbo, gpt4 and gpt4-turbo as the model in my request but nothing seems to make it give me back a larger response. Here is the core request that i send where the model is a variable based on a dropdown and the transcript is the prompt. Is there anything about this request that could be causing the poor response size?

			model: document.getElementById("gptModel").value,
			max_tokens: 4000,
			temperature: 0.7, 
			top_p: 0.9, 
			frequency_penalty: 0.5, 
			presence_penalty: 0.6, 
			messages: [
			  {
				"role": "user",
				"content": transcription
			  }
			]
2

ChatGPT wraps your conversation in another, hidden prompt. Maybe consider adding a system role message to instruct the model how to behave?

Also, do you know what you’re doing with your parameters? :thinking:

3

So for the messages array, add a new record at the start with role: system and then content along the lines of “please return a detailed response with X paragraphs per heading etc…” and then in the role: user content, give the standard prompt without needing to specify how verbose i want the response to be?

i asked chat gpt for the best parameter suggestions to get a longer output and this is what it gave me.

4

Yeah, the gpts don’t really know how to use themselves.

something like that, yeah.

Have you read through this? https://platform.openai.com/docs/guides/prompt-engineering

5

I’m also running into this issue and it’s frustrating. I don’t necessarily want longer responses, I just want it to have better responses.

Oftentimes I’ll get a really shorthand answer in the openai api response and then chatgpt gives this awesome bulleted or numbered detailed response that is exactly what I wanted.

I have no idea how to fix this…