External OCR API Integration

[savi8sant8s]

Question

Can an OCR API be run externally? To receive the output and add to the converter?

[dolfim-ibm]

It would be possible to implement it, but be aware of

1. The OCR latency will for sure increase substantially
2. This should be clearly communicated to the users opting for this type of OCR

[savi8sant8s]

@dolfim-ibm One more question. Is it possible to use only the document conversion part with an output like the one attached and skip the initial steps? If so, do any of the repositories (docling, docling-core, docling-parser) have a piece of code similar to this goal? I want to test doclingo to see its performance in transforming a standard OCR output into a more readable structure.

Output:

{
    "status": "succeeded",
    "createdDateTime": "2024-12-24T08:46:32Z",
    "lastUpdatedDateTime": "2024-12-24T08:46:34Z",
    "analyzeResult": {
        "version": "3.0.0",
        "readResults": [
            {
                "page": 1,
                "angle": 0,
                "width": 8.2639,
                "height": 11.6944,
                "unit": "inch",
                "lines": [
                    {
                        "boundingBox": [
                            0.7901,
                            0.9701,
                            1.0501,
                            0.9668,
                            1.0634,
                            1.1834,
                            0.8067,
                            1.1901
                        ],
                        "text": "Dunder Mifflin",
                        "words": [
                            {
                                "boundingBox": [
                                    0.7901,
                                    0.9701,
                                    1.0101,
                                    0.9668,
                                    1.0134,
                                    1.1868,
                                    0.7934,
                                    1.1901
                                ],
                                "text": "D",
                                "confidence": 0.981
                            }
                        ]
                    }
                ]
            }
        ]
    }
}

[dolfim-ibm]

Pointing you to the new tech report. In Figure 1 we outline the steps in the PDF pipeline: #574

If you have a scanned document (I assume since you refer to OCR) the PDF parsing is just identifying the image bitmap, which are then sent to the OCR engines.

In my opinion, the output you want is simply the DoclingDocument in result.document which has all the document components.

[savi8sant8s]

Thanks for information