Invoice Recognition Benchmark Comparison

CRS local tests · 200 invoices · 15 vendors, plus a 25-scan independent check

End-to-end structured recognition

PipelineVendorTotalInvoice #DateJobOverall
Gemma v0baseline96.094.074.097.046*83.4
Gemma v1resolver + Tulsa rules95.593.588.596.541.585.4
Gemma v2current proven stack97.093.591.596.544.286.8
Gemma 12B A/B92.586.584.595.063.385.5
Gemma E4B A/B93.572.586.097.524.577.6
Xberg replacementfull flattened text99.576.580.592.049.781.3
Xberg hybridtargeted evidence repairs10094.595.096.544.288.4

Percent correct. Strict total shown; corrected total is 96.0% for v2 and 97.0% for the hybrid when valid pre-tax subtotal matches are accepted. *Original v0 journal reported job as 40/46%.

Raw field evidence and speed

ExtractorInvoice #AmountVendorAll threeSeconds/document
Tesseract82.088.591.064.5
Xbergdigital PDFs93.095.087.085.00.013
Suryalocal GPU VLM OCR95.597.591.084.015.27

Independent image-only scan check

ExtractorInvoice #AmountVendorAll threeVerdict
Tesseract22/2525/2525/2522/25Works
Xberg default0/250/250/250/25No OCR
Decision: retain Tesseract for scans, use Xberg as a fast corroboration layer on digital PDFs, and reserve Surya for difficult layout-heavy fallbacks. The targeted Xberg hybrid is the strongest measured end-to-end pipeline.