Hi! I need advice on how to best approach AI enrichment. I have a company that I am enriching with AI as I don't have official company data. I create a prompt, I save it, but I run only on first row, it works. I run it on 10th row - it doesn't work and I update script. Finally, when script works for 10 rows, I run it for all rows, for example 50. But on row 27, the script doesn't provide me LinkedIn, although I manually checked that the LinkedIn link exists for that company. My question. What should I do? Update the script again and run on all rows again? Now I have a table of 50 rows but if there are 100 rows and part of script doesn't work on row 85? What is the best cost saving tip for this? Thanks! 🙏
It happens because AI enrichment can behave inconsistently across rows. the best cost saving approach is to iteratively test your prompt on a small sample first for example 5 10 rows with different types of data Once the prompt works reliably across that mixed sample then run it on the full table.
Daayem S. here is a Loom of my AI prompt https://www.loom.com/share/79e23faf5890405abd8c6f04c8fd585b
Okay so here's what I find what the LLM might be interpreting as a set of contradictory instructions in the screenshot. On the one hand we're telling it to state that no relevant information was found, and on the other hand we're telling it that if no verifiable information is found, say nothing - leave it as an empty string. LLMs have a really funny way of interpreting what we want them to do. Here's the 1st iteration of changes I'd try: In point #4, change it to: "- I preferred news from the last six to 12 months, including news about product launches. funding, etc. However, if no such news is found, return "No significant news found." For section 5, try changing the last line to: "- DO NOT fabricate news. If any of the information I asked you to find cannot be found, return a sensible statement such as "No significant news found". In any case, DO NOT RETURN an empty strong/column. Let's see if these work. If they don't, we can hop on a quick google meet and try to redo it together!
