Best way to use GPT for creating product descriptions

Hi there

We are a second-hand store and im figuring out ways to use the API (text-completion) to create product descriptions for a webshop.

This is my promt:

Complete / fill in, informations for a product description.

IMPORTANT: You must only respond if you have information about the entire product name, including both the brand and model. If you do not recognize the full product name, you must respond with “I have no information about this product”.

Now, please strictly follow these guidelines when completing this task:

  • Language: German
  • Lead Text: Compose two or three sentences describing the product, focusing on its primary use. Ensure that the text is written in the style of a reseller.
  • Markdown Syntax: Format the text using Trello Markdown syntax for optimal readability. Add two empty spaces at the end of each line to create a line break.
  • If you lack the necessary information to address a specific point, write “[no information]”.
  • Specifications: Include only the most essential product specifications.

Below an example of such a description:

GARMIN Fenix 6s

Der Fenix 6s ist ein leistungsstarke und robuste Sportuhr, das eine hervorragende Kombination aus Funktionalität und Design bietet. Mit einem 1,39 Zoll großen AMOLED-Display, innovativen Gesundheitsfunktionen und einer langlebigen Akkulaufzeit ist sie für Profisportler wie auch für den Alltag geeignet.

Brand:
Type:
Model:
Touch:
Release Date:

Sensors:

Functions:

Accesories:

Condition / signs of usage:

I send this as a promt + the name of another Product. For Example: GARMIN Strala 8800XT (which doesn’t exist).

The problem is, GPT made up facts when is has no knowledge about it. It ignores 75% of my rules. Lowering the temparature doesn’t make any difference.Should I switch to another endpoint other than “text-completion” or how can I overcome this issue?

Hi @CradleToCradle and welcome to the forum!

In your case the best solution would be to use fine-tuning. In the OpenAI docs your case is mentioned here explicitly. This is a bit more advanced then using Chat GPT, but if you want to achive a great result you might have to put the additional work in.


Fine-tuning - OpenAI API

If you want to stick to your approach and workflow you might also consider changing the presence penalty parameter to avoid getting this kind of response.
image