Document AI — three eras from OCR pipeline to VLM-native
Era 1: OCR pipeline
Tesseract / TrOCR detect
LayoutLMv3 layout
table recognizer
regex + domain rules
pros: cheap, deterministic
cons: brittle on new formats
Era 2: OCR-free specialists
Donut: image -> JSON
Nougat: paper -> LaTeX
DocLLM: layout-aware gen
swin / ViT encoder
pros: single model
cons: domain-specific
Era 3: VLM-native
Qwen2.5-VL native res
PaliGemma 2 doc-trained
Claude 4.7 at 2576px
frontier proprietary
pros: no pipeline
cons: cost + hallucination
benchmarks + 2026 recipe picker
benchmark
OCR+LLMv3
Nougat
PaliGemma 2
Claude 4.7
DocVQA
83.0
77.3
88.4
95.1
ChartQA
-
-
85.1
92.2
Math LaTeX
-
90.5
82.0
94.3
handwriting
65
-
80
92
2026 picker
10M invoices/day -> Era 1
scientific papers -> Nougat + VLM
mixed handwriting -> VLM-native
regulated -> hybrid cross-check
frontier gap
open 7B VLM: ~88 DocVQA
Claude 4.7: ~95, near-human