SOTA IDPby lasal.ai
Intelligent document processing · today

The same documents, today's models

The 63 invoices and purchase orders from the Legacy IDP, the same 11 fields and the same scoring. Seen: the documents the legacy rules were written on. Unseen: documents nobody wrote rules for. The full site (document viewer, how each pipeline works, live tries) is being built; this table is measured, not estimated.

pipelineruns onfields right · seenstraight through · seenfields right · unseenstraight through · unseentime per document
Qwen3-VL-8B: image to JSON
vision-language model
RunPod GPU91.5%29 / 4592.4%8 / 183.0 s
PaddleOCR-VL-1.6 reads, Qwen3-VL-8B extracts
document parser + LLM
RunPod GPU94.7%33 / 4591.9%9 / 185.1 s
Qwen3-VL-4B (4-bit): image to JSON
vision-language model
laptop GPU78.8%15 / 4583.3%8 / 1824.2 s
Qwen3-VL-2B: image to JSON
vision-language model
laptop GPU72.5%8 / 4577.3%5 / 1818.0 s
PP-OCRv6 (PaddleOCR) + the same rules
better OCR + rules
browser99.4%44 / 4555.6%3 / 1826.4 s
Legacy: TrOCR + 36 hand-written rules
OCR + rules
browser94.9%35 / 4555.1%3 / 18-
Legacy with a fine-tuned TrOCR + the same rules
OCR + rules
browser85.7%9 / 4551.0%0 / 18-
Built by lasal.ai · Qwen3-VL (Apache-2.0), PaddleOCR / PaddleOCR-VL (Apache-2.0), MinerU, GLM-OCR, TrOCR (MIT)