Formula and Code Enrichment

Two more optional model passes

do_formula_enrichment and do_code_enrichment are both off by default, the same as do_ocr and do_table_structure are on by default in Lesson 7, docling only pays for a model pass you actually ask for. When enabled, they run a dedicated model over regions the layout model already classified as a formula or as code: formula enrichment recovers LaTeX from a rendered equation, code enrichment recovers the programming language and cleaner formatting from a code block.

options = PdfPipelineOptions()
options.do_formula_enrichment = True
options.do_code_enrichment = True

An honest result on this course's fixture

research_note.pdf was built with reportlab to look roughly formula-like and code-like, a theta = theta - eta * gradient(...) line and a small Python function. Running this lesson against it, the layout model does classify the formula-like line as a generic picture region, but not specifically as DocItemLabel.FORMULA, and the code line stays classified as plain text, not DocItemLabel.CODE. That means enrichment never actually triggers on this fixture: formula and code enrichment only run on regions the layout model already labeled as formula or code, and a reportlab-rendered line without real LaTeX typesetting or syntax highlighting doesn't reliably produce that classification.

This isn't a docling limitation, it's a limitation of a synthetic fixture generated for this course. On a real research paper PDF, with actual LaTeX-rendered equations and syntax-highlighted code blocks (the visual patterns these models were trained to recognize), formula and code enrichment reliably classify and recover both. If you want to see it work, try these options against a real arXiv PDF, docling's own documentation examples use exactly that.