Introducing Kaori K1: High-Precision Spatial Document Synthesis and Multilingual OCR
- 120M parameter spatial cross-attention vision backbone
- Zero-shot layout reconstruction for multi-column PDFs, scientific formulas, and financial tables
- Native support for Devanagari, historical manuscripts, and Latin scripts with 99.4% structural accuracy
Today, Sorika Labs is announcing Kaori K1, our multimodal spatial vision engine engineered for zero-shot document hierarchy synthesis and high-precision OCR extraction.
The Challenge with Classical OCR
Overcoming bounding box drift in complex layouts.
Traditional document understanding pipelines rely on fragile, decoupled two-stage architectures: a bounding box detector followed by an optical text recognizer. In dense multi-column research papers, tables with nested cells, or historical manuscripts, these systems frequently scramble reading orders and lose hierarchical context.
Kaori K1 eliminates bounding box decoupling by formulating document perception as a unified spatial graph generation task. High-resolution pages are processed into continuous multi-scale patch tokens that preserve microscopic typographical geometry.
Developer API Integration
Single-line document extraction.
Developers can now query the Kaori K1 engine through our low-latency cloud endpoint or self-host the quantized model on local servers:
import sorika
client = sorika.Client(api_key="sk_live_...")
result = client.vision.parse(
document="financial_report_2026.pdf",
model="kaori-k1-spatial",
extract_tables=True,
language_mode="multilingual"
)
print(result.to_markdown())