Let me share my feedback and general feeling to the newly GA-ed feature of “Extract Rows”.
The level of accuracy and reliability is far low and the actual/real world use cases could not be satisfied as it does not read a relatively (a bit) complex template. During our testing, we saw error message like this after running Extract Rows task number of the times.
Error Message
Error encountered in step with name [New step]: Error: AppSheet AI Service was unavailable: Object reference not set to an instance of an object.
On the face of app, the user could not see error message, but the error is only accessible through the monitoring app which is also not user-friendly. User not sure what happened to the jobs after posting files to the apps.
Interestingly, I posted those files Extract Rows task failed onto the AI Studio/Gemini Chat app and simply asked AI “please extract XXX, YYY, ZZZ values and display as table. This works like a charm, the AI returned mostly accurate values out of input files. Very strange, why AppSheet AI repeatedly fails while we believe the same AI Engine is behind the scene.
Further testing showed the gaps between Extract/Extract Rows tasks. For those files failed with Extract Row task, then switched to existing “Extract” task to read the values. This BOT (task) did not returned the error at all and presented the extracted values with good reliability and level of the accuracy. Not sure why Extract Rows task mostly fails while Extract task works fine.
Overall, the usage of Extract Rows task seems to be pretty much limited, only satisfy the very much straight ward and simple story as use cases.