WorkBuddy extracts numbers from research reports for summaries, but units keep getting mixed up—anyone hit this?

This morning I came across a daily-ranking breakdown about the AI toolchain and data visualization space, saying that what everyone's competing on now is the "last mile" — whoever fills in the deployment step is the one with value. My first reaction after reading it was: isn't this exactly what I do every day?

My scenario is pretty unglamorous. After market close I need to dig out a few core numbers from a dozen or so research reports from the day — revenue, net profit, year-over-year — compile them into one table, then draw two trend lines. Doing it manually takes about forty minutes. I've been using WorkBuddy for two months, and before that I'd only used it to move tables around and split PDFs — basically treating it as a fancy Excel — so last week I figured I'd just hand this whole workflow over to it and see.

It didn't go well.

The table came out, but the units got mixed up. Two of the reports were in ten-thousands of yuan, the others in hundred-millions, and it just sorted by the raw numbers, so at a glance I thought some company's revenue had grown a hundredfold. Also, one report's table was an image, and the whole thing got skipped — the log seemed to show something like a parse failure, I don't remember the exact wording. The most annoying part was the data-source footnote line under each table — it got swallowed as a row of data, so the last column was all "Source: company announcement."

I tried a few things. First I fed the PDFs in one at a time, which was a bit better than dumping a dozen at once — at least I knew which one was causing the problem. Then I manually specified the column names in the field mapping, filled in both Chinese and English, and it worked for a while, but as soon as I switched to a new report it got messy again. I think there was an option related to units, or maybe I'm misremembering — anyway, I couldn't find it. Last weekend I wrote here about the daily-report date serialization issue, and later I just hardcoded the date format and it was fine — this is clearly a different kind of problem.

I did a quick check on the table it produced — accuracy is about seventy percent, and the other thirty percent needs manual fixing. The ROI isn't great right now; verification takes longer than doing it myself. But I think the direction is right, I just haven't found the correct way to use it yet.

Wanted to ask if anyone has used WorkBuddy for similar batch extraction. For things like units, do you have to hardcode them in the prompt, or is there another setting? Also, are image-format tables just not supported? Did I set something up wrong somewhere?