Extract AI task fails to select Ref value

AppSheet AppSheet Q&A
,
1

Hi

I want to use the Extract capability of AI task to fill out an Invoice table based on an invoice in pdf file. Among the various columns that I extract is one called [Supplier], which is a Ref column to the Supplier table. The [Supplier name] column in that Supplier table is the key and label for the table. The Supplier table has around 300 rows.

I started testing it (with both the AI automation test feature and the monitor console) and it works reasonably well, but for some invoices, the supplier name is properly extracted but not copied to the Invoice table, and I see a warning that there is no such row in the Supplier table. But there actually is and it is perfectly matched with the [Supplier name] entry (I double checked and they are indentical to the letter).

Has anyone experienced this as well? Are there limitations to Extract when dealing with a Ref column, like a maximum number of rows in the parent table?

Thanks

2

Hi , since I haven’t received any answer, I was wondering:

  • does anyone use AI tasks and the Extract function at all or am I the only one subscribing to a Enterprise Plus plan?? :sweat_smile:
  • if you do use it, no issues of that sort?

Thanks

G

3

It is difficult to opine without more precise details. When you say they are perfectly matched- do you mean visually or with formula?

If the names are matched visually, a remote possibility could be the AI extract introduces some invisible space/ special characters differently in different extracts. So you may want to ensure this is not the case by comparing two visually similar names in Invoice and Supplier tables. You could do it in a backend sheet for example.

4
5

I meant literally matched, because once it was extracted and I could see the supplier name in the automation monitor console, I copy/pasted it back into my supplier table. I ran the same automation again but it didn’t help…

Thanks anyway. I have asked tech support help on this and will report here what I learned from them.

G

6

Thanks Steve, but I didn’t find any hint at a possible explanation in those support pages.

Let’s see what tech support comes back with.

G

7

Dear @Steve and @Suvrutt_Gurjar

So here is what I eventually learned from tech support (and a warning in the automation monitor console): the AI extraction feature currently supports a maximum of 50 rows when trying to match data against a Ref column!

They suggested changing the column type to Text or Enum… I answered this was not an option because I need the parent-child relationship, and asked whether I could use an Enum type with a base Ref. To which they answered: “Enum with ref should have the same 50 row ref limit”. “Due to the limited context window we can pass to Gemini, we don’t want to give the ability to load too many rows so only the first 50 rows in the referenced table are loaded.”

So I’m afraid I will just have to wait for the limit to be increased because it will be useless for me otherwise… I just hope the wait won’t be too long: I guess they should be able to find some tricks to only pass a few columns to Gemini. And with Gemini’s 1 million token context window, it shouldn’t hold them back…

Cheers

G