DOCUMENT INGESTION PIPELINE

Embodiments manage data for document ingestion pipelines. Terms in a document may be determined based on a term location within the document. A layout graph that includes term nodes may be generated based on the terms such that each of the terms corresponds to a different term node. Relationships between the term nodes may be determined based on a traversal of the layout graph such that the layout graph may be partitioned into content shapes based on a strength of the relationships. The content shapes may be classified to reduce computational resources for extracting information from the document based on types associated with each content shape such that the types may be associated with extraction models that each support a plurality of different document formats. Extraction models may be employed to extract information from the classified content shapes.