Playground and API discrepancies

Hi team,
today I was brewing my first gpt-4 api based app in PHP and I am SHOCKED with discrepancies between responses from API and Playground.

Before I have started sending prompts/commands/content via api, I have prepared system and user prompts, verified temperature and top_p parameters and I was extremely happy with the results. The task was to rewrite article keeping core points, print it out in requested format “[Title:$title|Article:$content]”, use Markdown to mark every header (form h1 to h3). Results in Playground were PERFECT.

Thrilled, I have COPIED my Playground setup to API and… results are just AWFUL. Nothing is like in the Playground. Style is bad. Quality is poor. He is even quoting some of the system setup sentences! Of course format is nothing like expected - it is just a flow of words without general idea where does the tile ends and article starts. Headers are missing, so of course they are not Markdown-formatted as well… Just terrible experience.

So my questions is - am I doing something wrong? Am I missing some part of the setup, or maybe my expectations are too high? (but both Chat and Playground did wonderful job…)

Maybe share your exact API call?

Otherwise it’s simply not possible for anyone to even begin to offer meaningful insight.

Other things to add which would be helpful would be screenshots of the playground and ChatGPT interfaces where you got better results and a copy of the API generated output so we can see exactly what it is doing and how it is different from your expected results.

1 Like

OK, while trying to reproduce the issue, in English, it seems I have solved the issue.

Yep. You have to talk IN ENGLISH via API, even if you want the outcome in your native. Just add: “Deliver your output in xxx.”. So if you have similar problem like me - use English :slight_smile: