Player not loading? Watch on YouTube
Tobias compares OCR tools for document workflows and demonstrates an approval pipeline for invoices and contracts. He separates document intake, quality checks, text extraction, schema structuring, agent reasoning and follow-up actions. The pre-scan checks include sharpness, contrast, brightness and DPI before the heavier extraction step.
The demo uses Docling locally and Qwen 3 VL through OpenRouter. The speaker says the latter can also run on your own GPU, but this demonstration uses an API. LangExtract turns extracted text into defined fields, and a decision model checks those fields against rules in the code. Examples flag missing invoice quantities, a total that does not match subtotal plus tax, and a seven-page contract with automatic renewal but no notice window.
The comparison covers PaddleOCR, GLM-OCR and other open source document tools. Tobias reports that Docling works on a standard laptop and describes GLM-OCR as heavier. He recommends choosing by hardware, document volume and required precision rather than benchmark rank alone.
The failures matter for a local AI workflow: OpenRouter encounters issues during one test, while Docling responds locally but cannot extract useful information from a very blurry invoice. The tutorial ends with the speaker's guidance on choosing OCR for document search, complex legal documents and RAG.
