Advanced VLM backends¶
MinerU 2.5 Pro, PaddleOCR-VL 1.5, and GLM-OCR remain supported for Stage 1 benchmark extraction. Configure them through pipeline.strategy: vlm_ocr and the vlm YAML section.
These backends use separate environments and may start local inference servers. Their GPU/runtime constraints differ, so the repository does not ship a universal environment installer. Follow each upstream project's compatible installation instructions and either use the conventional local venv names listed in the uv guide or provide an executable explicitly:
| Backend | Python override |
|---|---|
| GLM-OCR vLLM server | GLM_VLLM_SERVER_PYTHON |
| PaddleOCR GenAI server | PADDLE_VLLM_SERVER_PYTHON |
They are intentionally excluded from the production quickstart.
Mathpix is represented separately with pipeline.strategy: mathpix_ocr. It
requires MATHPIX_APP_ID and MATHPIX_APP_KEY in .env. The typed runner
writes the normal Stage 1 flat output and publishes Markdown under
ocr-hints/<experiment>/, allowing a later sweep experiment to reference it
through runtime.ocr_hint_experiment.