Hi team -- I'm working on an AI enrichment, and I'm using field outputs with select lists. I've done this in the past, but for some reason the output is not respecting my select list options. Do you have any advice here?
It seems like you're having an issue with your AI enrichment output not respecting your select list options. To address this, let's first check a few things: 1. Make sure you've correctly defined your column outputs as "Select" data type in the AI integration setup. 2. Verify that your prompt is instructing the AI to use the specific options you've defined for the select list. 3. Double-check that the AI's response is being correctly mapped to the intended field. If you've confirmed these and are still experiencing issues, it would be helpful to see the exact JSON output. Could you please copy and paste the JSON into the chat so we can troubleshoot further?
To resolve the issue with your AI enrichment output not respecting select list options, let's check a few things: 1. Ensure you've defined your column outputs as "Select" data type in the AI integration setup. 2. Verify that your prompt is instructing the AI to use the specific options you've defined for the select list. 3. Double-check that the AI's response is being correctly mapped to the intended field. If you've confirmed these and are still experiencing issues, it would be helpful to see the exact JSON output. Could you please copy and paste the JSON into the chat so we can troubleshoot further?
Hey Andrew, I'm not working for Clay, but can help out. Can you share an example of the prompt that you're using for the output?
Hey -- Part of the prompt includes this: "**Location_Rating:** This determines whether or not a product is safe to operate in humid, moist, or wet locations. These types are “Dry Location”, “Damp Location”, “Wet Location” and the output can only be one of these three options." I have other "select" type outputs, and it's working fine. I'm just getting an error with this one...
Hey there Andrew thanks for reaching out and appreciate you jumping in David! Taking a look at your table I see that it the column is now outputting the correct result here, did you still need any help with this.
Thanks LuisArturo -- I ended up adjusting the prompt, and I am getting more reliable output. I've just never seen an instance where an AI action in Clay didn't respect the select list options (regardless of the prompt). Is this something you've seen before?
Interestingly, I'm seeing this with another attribute as well.
Hey Andrew, Could you show us where you're seeing this issue as well, please?
Sure thing. I'll run some more tests and try it again.
Rows 22, 24, and 25 are good examples of the error.
Row 22 has a Location_Rating of "Wet", but in the output settings for the AI column, I have that output restricted to 3 values. It is also addressed in the prompt.
Hey there Andrew thanks for sharing this, I have seen this occur before, but that is when there is a large number for the number outputs expected, as this puts a lot of strain on the AI to perform. Granted it usually only happens for a handful of rows just like you have experienced. Sometimes reworking the prompt can help it get the right answers, other times just rerunning the rows can get the correct response. This is essentially coming down to the AI model being pushed to its limits here.
Gotcha. Thanks for clarifying. What do you think is the upper limit of reliable output? I'm wondering if it's worth breaking this into a couple of prompts. If I have to break it into 10 prompts for reliable output, that starts to impact processing cost a bit more.
Also, it the limitation based on the number of outputs or the overall complexity of the prompt? In other words, if I pass the exact content and url I want research done with so there are fewer research steps, would that have any impact?
And finally, does the model matter here? I was using 4o-mini to keep costs low, but if I need to consider a different model, that would be interesting (also a cost impact).
Hey Andrew sorry for the delay here, you could break into multiple prompts here to help ease the strain on the AI, however that would probably require 5-6 separate prompts to do so. Generally, I would say around 10 outputs is the limit, however my concern here is that it would require about 5 or 6 AI columns to perform this as mentioned before, which would just eat up too many credits at once. A higher model would see a better performance as well, as it would have stronger processing power in order to read and perform the prompt you have given it.
Okay. Thanks for the insights on this. I'm going to think about it more and see if I can reduce the number of outputs and get creative with some things. 🙂
.png){kind=logo}
