Moshi developed an in-house vision language model to extract data from unstructured PDF documents