Design proposal for remote apis

[dolfim-ibm]

Currently, Docling is designed to run only local models which are not sharing any user data with remote services. We see the need to allow new generation of models to communicate with remote services. Among others, possible use cases are invoking OCR engines from cloud vendors and the usage hosted LLMs.

The following design is proposing to add explicit options for the users to opt-in in remote communication. All PRs will be reviewed by the maintainers to make sure this is consistently propagated in future changes.

Proposal

Settings for remote services

class PipelineOptions(BaseModel):
    """Base pipeline options."""

    allow_remote_services: bool = False

Then the value will be propagated to the init of the models which would like to invoke remote services.

• If the value is not True, this raises an exception

Usage

pipeline_options = PdfPipelineOptions(allow_remote_services=True)

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

[simonschoe]

@dolfim-ibm great idea. It would be super helpful to also have well-documented examples of how the user could plug-in a self-hosted OCR model via API and the required output format to be further processed by the pipeline