kreuzberg-candle-ocr 5.0.0-rc.35

Candle-based VLM OCR engines for Kreuzberg - pure-Rust transformer OCR (TrOCR, PaddleOCR-VL, GLM-OCR)
docs.rs failed to build kreuzberg-candle-ocr-5.0.0-rc.35
Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.

kreuzberg-candle-ocr

Candle-based VLM OCR engines for Kreuzberg. Pure-Rust transformer OCR via candle.

Supported models (per-model sub-features):

  • trocr — Microsoft TrOCR (printed/handwritten, ~330M, MIT). Line-level only: TrOCR expects a single text line per image and produces poor output on full-page documents. Pair with a text-detection stage that crops text regions before recognition.
  • paddleocr-vl — PaddleOCR-VL (0.9B, Apache-2.0, multi-task: OCR/tables/formulas/charts). Full-page vision-language model, emits markdown directly.

Device pass-through features mirror candle's own: cuda, metal, mkl, accelerate.

When depending on kreuzberg, the equivalent aggregate features are candle-cuda, candle-metal, candle-mkl, candle-accelerate. Enable alongside any candle-* backend — without one of them the candle build remains CPU-only, and PaddleOCR-VL CPU decode is impractically slow.