Enrichment Pipeline#

The enrichment pipeline is Docling's post-document-assembly processing phase. After the main layout analysis, OCR, and table extraction pass assembles a DoclingDocument, an ordered list of enrichment models (self.enrichment_pipe) runs against the assembled document to annotate specific element types with additional information. All built-in enrichment models are disabled by default because they require additional model inference and increase processing time. 1

Execution Loop#

The loop in BasePipeline._enrich_document() iterates each model in enrichment_pipe, uses the model's prepare_element() to filter and prepare candidates, chunks them into batches via elements_batch_size, and calls the model as a callable on each batch. The inner iterator must always be fully exhausted. 2

Built-in Enrichments#

3

Pipeline Composition#

ConvertPipeline initializes the base enrichment_pipe with the picture classifier, picture description, and chart extraction models. StandardPdfPipeline._init_models() prepends CodeFormulaVlmModel to the inherited list using spread syntax — so code/formula enrichment always runs first. 4

Base Classes#

All enrichment models in docling/models/base_model.py extend one of two abstract base classes:

BaseEnrichmentModel 5#

Simplest variant. prepare_element() returns the element as-is if is_processable() returns True, otherwise None. No image cropping. Used for text-only enrichment.

BaseItemAndImageEnrichmentModel 6#

Used when the enrichment needs a cropped image of the element's bounding box. prepare_element():

• If the element is a PictureItem with an embedded image, returns it directly.
• Otherwise, crops the region from the source page image at self.images_scale with optional expansion_factor bounding-box expansion. 7

Returns an ItemAndImageEnrichmentElement (a (item, image) pair) ready for batch inference. 8

Both classes inherit from GenericEnrichmentModel[T], which declares the three required abstract methods: is_processable, prepare_element, and __call__. 9

Plugin Architecture#

Custom enrichment models can be injected in two ways:

1. Subclass the pipeline directly (recommended for development)#

Subclass StandardPdfPipeline, override __init__, and replace self.enrichment_pipe with your model list. Pass the custom pipeline class via PdfFormatOption(pipeline_cls=...). 10

class MyPipeline(StandardPdfPipeline):
    def __init__(self, pipeline_options):
        super().__init__(pipeline_options)
        self.enrichment_pipe = [MyEnrichmentModel(...)]

2. Setuptools entry-point plugins#

The factory system uses a PluginManager (backed by pluggy) to discover registered models from installed packages via setuptools entry points. External plugins are only loaded when allow_external_plugins=True is set in PipelineOptions — by default only modules whose names start with "docling." are trusted. 11

Production Implementations#

Key Files#