2 Comments
User's avatar
Yair Itmar's avatar

Great project!

Did you try Argon?

You can also create a custom OCR using weighted scores optimized for the each text-type, and then have low-confidence areas adjudicated by vlm.

I've had success with that workflow.

Ezra Brand's avatar

Hey, glad you like the project!

I haven't tried Argon directly. I'm using Claude opus 5.5 as orchestrator, and it tests various models using Openrouter (as I describe), including Google ones. And Claude essentially does what you're describing