Issue 115983 states "Users are noting that Auto Invoicing isn't selecting data as well as previously. This includes data being transposed, dropping digits, switching digits, and inputting 5’s instead of an “S”."
This needs to be fixed NOW - defeats the whole purpose of using this process if you have to manually verify every character for accuracy.
Laura Moran
| Company | Kolb Grading |
| I need it... | Yesterday...Come on already |
Dear Viewpoint Suggestion Box contributor;
We at Viewpoint sincerely thank you for your contribution to Suggestion Box on how we can improve Viewpoint products. While we can’t do everything at once, we rely upon your feedback to help guide the prioritization of our product improvements, and Suggestion Box is a critical tool for us to understand and prioritize our customers’ needs.
Viewpoint reviews Suggestion Box regularly for all of our products and updates statuses, adds comments, and performs various house-keeping (including deleting) as needed to ensure that Suggestion Box is maintained as a productive environment for product enhancements requests.
© 2023 Trimble Inc. All Rights Reserved. Viewpoint®, Vista™, Spectrum®, ProContractor™, Jobpac Connect™, Viewpoint Team™, Viewpoint Analytics™, Viewpoint Field View™, Viewpoint Estimating™, Viewpoint For Projects™, Viewpoint HR Management™, Viewpoint Field Management™, Viewpoint Financial Controls™, Vista Field Service™, Spectrum Service Tech™, ViewpointOne™, ProjectSight® and Trimble Construction One™ are trademarks or registered trademarks of Trimble Inc. or its affiliates in the United States and other countries. Other names and brands may be claimed as the property of others.
And for clarity and those finding this post later, the GPT-4o model is still in place, but running with much improved OCR capabilities.
The release shipped last night, and according to our internal metrics accuracy significantly improved for invoice total, date & number. Vendor was a more moderate improvement but we have some ideas to work on there.
Thank you Sam for the update!
Thanks for raising this - we are actively working to improve the model accuracy and I totally understand how it's frustrating to have to squint at every character to catch close-but-still-incorrect answers. Full details/timeline of events below:
The GPT-4o model is much more capable than the original model, but had a couple of issues immediately after release (during the week of 9/21) where it was unable to handle the volume and performed really poorly for most of that week. We spent most of that week initially triaging issues and stabilizing performance.
After we got the initial issues stabilized we've been troubleshooting why accuracy hasn't improved like we expected. The root cause seems to be the built-in OCR on the new model isn't as good as what we were using previously, so the "smarter" model isn't giving good results because it can't distinguish similar-looking characters like S from 8 from 5, 4 from q, 1 from l etc. We're working to implement the better OCR alongside the better model. That should be ready for release this month (Oct 24th, 2 weeks away). It's frustrating that this didn't show up in our initial testing, but I'm working to ensure we are testing with more realistic data going forward. The new update should bring accuracy above what it was with the old model.