Have you ran your pipeline against an open benchmark like https://github.com/opendatalab/OmniDocBench?