Investigating Claygent's Confidence Levels Based on Prompts

Is Claygent’s confidence influenced by the prompt itself? I get the impression that it actually does a great job for most rows here but somehow gives a low confidence. Worth looking into it?

11 comments

Channeled
APP
·
·
Yes, Claygent's confidence is directly influenced by the quality and structure of your prompts. The effectiveness of Claygent heavily relies on three key elements: 1. Clear context setting at the beginning of your prompt 2. Step-by-step instructions that prevent incorrect assumptions 3. Explicit output specifications including formatting requirementsWhen you get low confidence despite good results, this often relates to how well your prompt aligns with the model's strengths. For instance, our Neon model excels at structured data extraction and answer formatting but may show lower confidence when dealing with complex reasoning tasks.To improve confidence scores while maintaining good performance, try: ** Being more specific about the data points you want to extract ** Structuring your prompts with clear, sequential instructions * Using the right model for your task (Neon for data extraction, GPT-4/Claude for complex reasoning) Sources: Prompt Engineering Crash Course [Clay 101] - Clay University Claygent: AI Web Scraper [Limitless Research] - Clay University Prompt Engineering Crash Course [Limitless Research] - Clay University
Andreas W.
·
·
influenced meaning if the prompt asks for verification steps, and it finds like other sources, does that lead to low confidence? Column “AI Innovation Labs” https://app.clay.com/workspaces/439843/workbooks/wb_4nNiN9KKeKAh/tables/t_b43X7rtpWEtF/views/gv_WpTXxYzKc74r
Channeled
APP
·
·
Looks like you're still looking for help. We're escalating this over to our support team and someone should be in touch soon. In the meantime, feel free to include any other information that may be helpful in solving your issue!
Tanvi R.
·
·
Hi Andreas, thanks for reaching out! Happy to take a look here for you
Andreas W.
·
·
awesome. No biggie but it’s a bit off. It looks like the ones that have a ‘false’ as response are marked low, by and large.
Andreas W.
·
·
for that would mean “claygent didn’t find enough / good information” but it did pretty good actually
Andreas W.
·
·
renamed it “clay support F1: AI Innovation Labs”
Andreas W.
·
·
I’m bumping this Claygent to Neon, much better for the task. I think I just overloaded 4o mini with this one. Good learning, but I’m still not fully understanding ‘confidence’
Tanvi R.
·
·
Hey Andreas, the confidence score essentially relates to how confident Claygent is in its response completely answering your prompt. Green values show Claygent was able to find the most information related to answering your prompt, where red means it wasn't able to fully answer your prompt with the information it found (even if the research was thorough)
Tanvi R.
·
·
Let me know if this makes sense and if you have other questions!
Channeled
APP
·
·
Hi Andreas W.! This thread was recently closed by our Support team. If you have a moment, please share your feedback:

Investigating Claygent's Confidence Levels Based on Prompts | Clay