I've been trying to whittle down the number of failed rows. I made another "Use AI" column named "Web Search Error" that analyzed the Reasoning output from the Tax Prep US Status Use AI column and parsed if the web search failed or not. After filtering to the failed ones, I've re-ran the main LLM call and repeated a few times. Have now gone from 2.4k failed rows to only 129.
I've ran nearly identical prompts in the past though with nowhere near this high of a failure rate, so I feel like something funky is going on