docling-surya-ocr

Docling OCR plugin backed by Surya. It provides the Docling OCR engine key suryaocr and a stable Python import namespace docling_surya_ocr.

The repository currently contains three runtime paths:

• surya2: target production path, using Surya 2 full-page OCR with llama.cpp on CPU by default and vLLM as a supported GPU backend.
• surya1: validated v1-style Surya recognizer path, kept as an explicit rollback and benchmark reference.
• unlimited: experimental comparison backend for internal evaluation only; it is explicit-only and not part of the Surya production path.

For UVO-style document processing, the intended production shape is Docling + Surya 2 + llama.cpp on a large CPU host, with page-level parallelism tuned by benchmark. GPU/vLLM stays available behind the same backend contract so the converter does not need to be rewritten later.

Install

From PyPI after release:

pip install docling-surya-ocr

From a built wheel:

pip install docling_surya_ocr-0.1.0-py3-none-any.whl

For local development:

pip install -e .[dev]

Runtime dependencies are declared for Python >=3.12,<3.13:

pip install "docling>=2.115,<3" surya-ocr

For GPU validation, install a CUDA-enabled Torch build that matches the target host/container.

Docling Usage

Use Docling's external OCR plugin seam and enable external plugins:

from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions
from docling.document_converter import DocumentConverter, PdfFormatOption
from docling_surya_ocr import SuryaOcrOptions

pipeline_options = PdfPipelineOptions(
    do_ocr=True,
    ocr_model="suryaocr",
    allow_external_plugins=True,
    ocr_options=SuryaOcrOptions(
        lang=["sk", "en"],
        backend="surya2",
        surya2_inference_backend="llamacpp",
        surya2_inference_parallel=4,
    ),
)

converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options),
    }
)

result = converter.convert("document.pdf")
text = result.document.export_to_text()

The package also exposes a compatibility import alias for deployments that already use the upstream plugin import path:

from docling_surya import SuryaOcrOptions

The Docling OCR model key remains suryaocr.

Backend Selection

Select the plugin backend in code or through SURYA_OCR_BACKEND:

SuryaOcrOptions(backend="auto")    # prefers Surya 2 when available
SuryaOcrOptions(backend="surya2")  # target Surya 2 path; fails closed if unavailable
SuryaOcrOptions(backend="surya1")  # validated v1-style rollback path

export SURYA_OCR_BACKEND=surya2

auto prefers Surya 2 when the Surya 2 API is available. If Surya 2 is unavailable and no strong Surya 2 runtime setting is present, auto can fall back to surya1. Use explicit backend="surya2" for production validation when fallback must not hide failures.

The Surya 2 inference backend defaults to llamacpp:

export SURYA_OCR_BACKEND=surya2
export SURYA_INFERENCE_BACKEND=llamacpp
export SURYA_INFERENCE_PARALLEL=4

vLLM is supported as the GPU inference backend:

export SURYA_OCR_BACKEND=surya2
export SURYA_INFERENCE_BACKEND=vllm
export SURYA_INFERENCE_URL=http://localhost:8000/v1
export SURYA_INFERENCE_PARALLEL=8

CPU Production Path

For Surya 2 CPU production, use llama.cpp and tune page-level parallelism:

export SURYA_OCR_BACKEND=surya2
export SURYA_INFERENCE_BACKEND=llamacpp
export SURYA_INFERENCE_PARALLEL=4
export SURYA_DEVICE=cpu
export SURYA_CPU_THREADS=8

Benchmark SURYA_INFERENCE_PARALLEL=1/4/8/16/24/32 on the target EPYC host before selecting a default. The optimum may arrive before all CPU cores are saturated because Surya 2 OCR can become memory-bandwidth or decode limited.

GPU Validation Path

For Surya 2 GPU attach mode, run or attach to an OpenAI-compatible vLLM Surya inference server and configure:

export SURYA_OCR_BACKEND=surya2
export SURYA_INFERENCE_BACKEND=vllm
export SURYA_INFERENCE_URL=http://localhost:8000/v1
export SURYA_INFERENCE_PARALLEL=8
export SURYA_DEVICE=cuda
export SURYA_PRECISION=fp16

Verify CUDA from inside the runtime container:

python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"

Options

SuryaOcrOptions can be configured directly or via environment variables.

Option	Environment variable	Example
Languages	SURYA_OCR_LANGUAGES	sk,en
Backend	SURYA_OCR_BACKEND	auto, surya1, surya2
Device	SURYA_DEVICE	cpu, cuda, auto
GPU devices	SURYA_GPU_DEVICES	0
GPU memory fraction	SURYA_GPU_MEMORY_FRACTION	0.8
Precision	SURYA_PRECISION	fp16, fp32, bf16, auto
CPU threads	SURYA_CPU_THREADS	8
Recognition batch	SURYA_RECOGNITION_BATCH	16
Detection batch	SURYA_DETECTION_BATCH	16
Docling OCR batch	DOCLING_OCR_BATCH	4
Preserve provenance	SURYA_OCR_PRESERVE_PROVENANCE	true
Surya 2 inference backend	SURYA_INFERENCE_BACKEND	llamacpp, vllm, auto
Surya 2 inference URL	SURYA_INFERENCE_URL	http://localhost:8000/v1
Surya 2 keep server	SURYA_INFERENCE_KEEP_ALIVE	true, false
Surya 2 parallelism	SURYA_INFERENCE_PARALLEL	8

The codebase includes an explicit-only unlimited backend for internal experiments. It is not selected by auto and is intentionally omitted from the main Surya production configuration above.

Smoke Test

After installing the wheel, run a conversion without PYTHONPATH=src:

python scripts/smoke_convert.py sample.pdf --plugin suryaocr

Force the Surya 2 CPU path during smoke validation:

python scripts/smoke_convert.py sample.pdf \
  --plugin suryaocr \
  --backend surya2 \
  --inference-backend llamacpp \
  --inference-parallel 4

The smoke test fails closed if Docling external plugins are disabled, the plugin key is unavailable, Docling logs an OCR-stage failure, no text is produced, or Docling page/bbox provenance is missing.

Benchmark

Corpus benchmark example for the Surya 2 CPU path:

python scripts/benchmark_docling_corpus.py benchmarks/pdf-corpus \
  --plugin suryaocr \
  --backend surya2 \
  --inference-backend llamacpp \
  --inference-parallel 16 \
  --jsonl benchmarks/results/surya2-llamacpp.jsonl

The historical benchmark report for this repository is in:

benchmarks/results/docling-surya-ocr-benchmark-report.md

Those historical measurements are for the v1-style backend. They are retained as a rollback/reference baseline, not as Surya 2 llama.cpp production evidence.

Surya 2 Model License

Surya code is Apache 2.0, but Surya 2 model weights are distributed under a modified OpenRAIL-M license. Review the weight license before commercial/production rollout; this package does not vendor or redistribute model weights.

Compatibility Notes

The target verified runtime is Docling 2.115.0. In that version:

• BaseOcrModel.post_process_cells uses the docling_2115 style and may mutate the page in place.
• BoundingRectangle is available from docling_core.types.doc.page.
• Docling provenance uses item.prov entries with page_no and bbox.
• Surya v1 recognition is available through surya.recognition.RecognitionPredictor; Surya 2 additionally expects surya.inference.SuryaInferenceManager.

Development Checks

python -m pytest tests/unit tests/contract
python -m ruff check src tests scripts
python -m mypy src