Skip to main content

The term "AI" is now widely used to make our tools seem more powerful. While I recognize the benefits of integrating AI into SF, it does not yet seem to be functioning stably in all areas. This could be due to a combination of insufficient documentation, lack of validation, or limited personal experience.

Here are two examples:

Create a sample data model with AI

I recorded the description in Dutch (if this needs to be in English, please mention it in the instructions). The result after performing this enrichment is: "ERROR: The data model is not valid JSON." This error message is meaningless to me. Is there another way to zoom in on what is going wrong?

 

ERROR: The data model is not valid JSON.

Add comments to code with AI

As a developer, I currently have no idea what will be changed apart from the comment lines being added or revised in the template. It would be helpful if this screen already showed what is going to be changed by means of a comparison. Currently, you can only guess and compare different templates afterwards.
 

Add d.?.] comment lines to template e...]

I don't want to forget to mention that there are certainly already some great working applications that have been released.

Hi Dennis,

For the first bit, we do rely on OpenAI returning a valid JSON to work with. Our Support team uses this feature a lot and we haven't encountered the error you are receiving. If you try the prompt in English, does it perform better? Which OpenAI model are you using? The newest models work the best from our experience. Feel free to create a ticket for this and add the used prompt so we can take a look to get it even more robust.

All of the code that is run is available for you to see in the Maintenance tab, then find the AI completion query. The prompt you enter along with some guidance prompt will be send to OpenAI. What is returned, the completion, is analysed and the data model is extracted from the JSON. If OpenAI returns an invalid JSON, that is faulty on their part but then it's up to us to find a suitable solution to it.

For the second bit, we have on our backlog some improvements to indeed be able to see that the model update exactly will perform. For the initial release it was left out due to complexity and time constraint reasons. We will be looking into how to make this user friendly and developer friendly as well.


Running it with an English description does give the desired result. Of course, there are countless ways to handle this. From including it in the instructions, giving a notification, to baking in a step that automatically translates the input to English before sending it to the AI provider.
 

OpenAI model gpt-4-turbo

 

OpenAI model gpt-4-turbo

 Looking forward to upcoming developments.