Reconstructing a PDF's logical structure

PDF is a presentation format: it specifies glyphs and their coordinates on a page, without encoding logical structure. Nothing in it distinguishes a heading from body text, delimits a table, or establishes reading order. Naive extraction concatenates glyphs in stream order, which interleaves the columns of a multi-column layout, linearises a table into a sequence of cells stripped of its headers, and loses section hierarchy. The degradation is silent: the extracted text remains readable while having lost the relations that carried its meaning.

The IBM Research team described in 2024, with Docling, a pipeline that treats layout analysis as a vision problem. An object detection model segments the page into typed regions (heading, paragraph, table, figure, header) whose reading order is then reconstructed; a second, specialised model recovers table structure, including merged cells and hierarchical headers. An optical character recognition engine takes over on pages with no text layer. The output is a structured representation, not a character stream.

This is the first step of indexing in DeepView, and it conditions everything that follows: a row-column relation lost at this stage is unrecoverable, whatever the quality of later steps. Passage segmentation follows this structure rather than a fixed length, which avoids splitting a table or separating a heading from the section it introduces. The whole pipeline runs on your hardware, scanned documents included, with no outbound network access.