Interactive document parsing demo
One document. Four OCR profiles.
Our parsing pipeline using different OCR models. Each output identifies its OCR model; generated visual descriptions are labeled Gemma.
70-page evaluation · ranking and metrics ↗
Inspect evaluation references · source pages + annotations ↗
PDF + parsed output
Page navigation keeps both panes in sync.
View settings
Run details & metrics
NVIDIA GeForce GTX 1660 Ti · 6 GB VRAM. Visual enrichment uses remote Ollama Cloud.
Accepted pages—Full saved selection; not just demo excerpts
Median page time—
GPU memory—
This page—
Timing and memory scopes differ across backends.
Secondary downstream study
When does extra RAG complexity help?
The completed comparison uses saved parsed documents and the same 100 questions. It evaluates chunking, retrieval and answering—not parsing-model accuracy.
| Approach | Strict judge passes | Citation support | Scored citations | Meaning |
|---|