kreuzberg-dev/kreuzberg v4.9.5

on GitHub

Fixed

• #790: Fix GPU acceleration — kreuzberg now bundles CPU-only ONNX Runtime by default (zero-config). When a GPU execution provider (cuda, tensorrt, coreml) is explicitly requested via AccelerationConfig but unavailable, kreuzberg returns an error with setup instructions instead of silently falling back to CPU. Auto mode gracefully falls back to CPU with an info log. For GPU support, set ORT_DYLIB_PATH to a GPU-enabled ONNX Runtime.
• #791: Fix DOCX OCR extraction — OCR now runs on embedded images before document rendering, and OCR text is injected into the rendered output. Previously, OCR results were discarded and replaced with placeholder text.
• #783: PaddleOCR backend not utilizing GPU (CUDA) despite AccelerationConfig — AccelerationConfig from ExtractionConfig was never reaching PaddleOCR ONNX sessions, silently falling back to CPU. Acceleration is now propagated through OcrConfig to all OCR call sites (image extractor, PDF OCR).
• #779: Expose PaddleOcrConfig in Python bindings and update OcrConfig for backward compatibility.
• #792: Fix Ruby gem packaging — exclude staged libpdfium.dylib from gem artifacts by narrowing the native extension glob to only include the compiled kreuzberg_rb.* extension.

Added

• GPU CI workflow (ci-gpu.yaml) targeting self-hosted GPU runners with NVIDIA GPUs.
• Comprehensive GPU integration tests covering all ORT-accelerated paths: PaddleOCR (det/cls/rec), layout detection (RT-DETR), embeddings, document orientation detection, and end-to-end extraction.