I have a simple scenario that is designed to use ChatGPT to rewrite Press Releases into an informational post. (Basically: Airtable - ChatGPT - Airtable)
The Make scenario pulls the link to the Press Release and sends it to ChatGPT to render and then the output is updated back into Airtable.
The issue that I run across is that the ChatGPT render refuses to keep the quote and author the same and hallucinates a new answer. I have a solid prompt that when I use it in any LLM (including ChatGPT) it will output correctly but when I run the same prompt through a make.com scenario it hallucinates.
My settings in the ChatGPT module are
Message 1 - Role: System: [ I have my prompt here that pulls in the url from aitable and gives the prompt instructions]
Message 2 - Role: User: [ Button: URL. *This is assigned to a button in airtable that activates the scenario.]
Max Tokens: 0
Temp: .5 [This is to keep the verbiage similar to the Press Release.]
Not sure why it hallucinates every time inside of make when it works flawlessly directly inside ChatGPT. Would it be the System or User role that I am making a mistake on?
There is a final output so the scenario runs it is just that it isn’t following instructions inside ChatGPT when I use the prompt inside of Make.
Any thoughts or advice would be greatly appreciate it.
One possible reason it might consistently make up something and ignore your input is if it’s not actually getting your input; or similarly if the net set of messages it is receiving for the chat is confusing.
Idea: install the Make DevTools and watch the API traffic going to OpenAI and confirm that the system/user/assistant messages look right and correctly contain the inputs you think you’re passing.
Also, confirm that the same messages work as expected outside of Make within the OpenAI playground
To allow others to assist you with your scenario, please provide the following:
1. Relevant Screenshots
Please share screenshots of your scenario, any error messages, relevant module fields, and filters in question? It would really help other community members to see what you’re looking at.
You can upload images here using the Upload icon in the text editor:
2. Scenario Blueprint
Please export the scenario blueprint file to allow others to view the mapped variables in the module fields. At the bottom of the scenario editor, you can click on the three dots to find the Export Blueprint menu item.
3. Output Bundles of Modules
Please provide the output bundles of the modules by running the scenario (or get from the scenario History tab), then click the white speech bubble on the top-right of each module and select “Download input/output bundles”.
A. Upload as Text File
Save each bundle contents in your text editor as a bundle.txt file, and upload it here into this discussion thread.
B. Insert as Formatted Code Block
If you are unable to upload files on this forum, alternatively you can paste the formatted bundles.
These are the two ways to format text so that it won’t be modified by the forum:
Method 1: Type code block manually
Add three backticks ``` before and after the content/bundle, like this:
``` content goes here
```
Method 2. Highlight and click the format button in the editor
Providing the input/output bundles will allow others to replicate what is going on in the scenario even if they do not use the external service.
Following these steps will allow others to assist you here. Thanks!
Thank you for the advice. That helped me troubleshoot. Ultimately, it came down to the ChatGPT API. I need to address the differences between the prompt structure when using the API specifically. But using the Make DevTools really helped go step by step to debug my issue.
