See screenshot. I have fed my assistant a json file containing recent blog posts about food safety issues. In my run instructions, I tell it to note any articles it references, providing their title and url. Obviously I am including the title and url of each article in the json file.
Annoying and weird: It does not abide by my instruction to cite the referenced articles until I demand it do so in a user message. It ignores the run message.
The following prompt has been helpful to me whenever I see unexpected behavior from an Assistant. In your example it would go like this:
Prompt: “Using the content in fsn.json…”
Response: [Response without references]
Prompt: “You did not add the title and link of each referenced article in your response even though I asked you to do so. How would you change the description of the assistant or tools to avoid this problem in the future”
Response: [Hopefuly! Useful recommendation on how to fix the issue]
Try it out and let us know the model tells you!
It tells to change my workflow so that after each response, I ask it to cite it’s sources as a separate message.
That’s pretty obnoxious and makes me want to throw my macbook out the window, but fortunately I am getting much better results with the following run instruction:
Using the content in fsn.json as reference material, along with the legal expertise of a seasoned attorney, respond to the user messages.
If you use any of the articles in fsn.json, add a list to the end of your response, noting the title and link for each referenced article. This is very important. Do not ignore this. Do it in this exact format at the end of your response:
- Post Title Here | example.com/example-post
- Other Post Title Here | example.com/example-post-2
Agree, it can be stubborn sometimes. Reply that you do not want to change the workflow and would like it to either tell you how to change the description, or recommend the implementation of a function to help with this functionality.