Route PDF conversions through analysis-driven extraction

Routes every PDF-to-another-format conversion through a short classification step before choosing tools or CLI flags, instead of always running one fixed pipeline. It classifies the source PDF (native digital, OCR-noisy, image-only/scanned, slide deck, medical/lab report, table-heavy business/finance document, narrative report, or mixed-layout document), classifies the desired output shape (markdown-with-html for fidelity-critical cases like tables, medical reports, slides, and mixed layouts; plain markdown when clean text matters more than layout; html when visual structure matters more than LLM readability; text for quick linear extraction; json for downstream machine processing; docx for editable office output), then chooses the strongest extraction route for that combination and validates the result on representative sections - not just the top of the file - before delivering it, retrying with better settings if the first pass is wrong. opendataloader-pdf is treated as the default, primary conversion engine for every conversion task; other tools (OCR, pdfinfo/pdftotext -layout, manual repair) are used only to classify, validate, preprocess, or clean up, never as the default replacement route. Once output exists, it runs a post-conversion repair pass split into three buckets - cleanup (noise reduction without changing meaning: footers, page numbers, stray symbols, empty rows), structural correction (fixing attachment problems where content was found but structure was wrong: flattened tables, fused columns, notes merged into result rows), and route retry (when the extraction path itself was wrong) - following a fixed retry order: same engine with better flags, then a different output shape, then hybrid/full mode, then cleanup/repair, then OCR preprocessing, and only as a last resort a non-OpenDataLoader fallback.

Use it whenever a user wants a PDF converted to .md, .html, .txt, .json, .docx, or structured notes, especially when the source may be scanned, OCR-heavy, table-heavy, slide-based, medical, academic, or multi-column. Never accept a conversion just because the first page looks correct - always validate at least one early section and one structurally difficult section (a real lab table for medical PDFs, a dense diagram or pseudo-table for slide decks), and treat a good command exit code with visibly poor structure, flattened table rows, or labels detached from their values as red flags that the output is not ready. Document fidelity (correct content, attachment, and section structure) is prioritized over visual fidelity (matching the original page layout); do not sacrifice semantic correctness to imitate the original page visually. When changing opendataloader-pdf's own behavior rather than just running a conversion, validate against the real target PDF plus at least one difficult benchmark case, and check that a fix for one document class does not reopen a regression in another.

Input is a source PDF plus a target format. Fast classification checks:

pdfinfo input.pdf
pdftotext -layout input.pdf -

Default document-specific routes:

opendataloader-pdf -f markdown-with-html --table-method cluster --image-output off

for medical/lab PDFs and table-heavy documents (verifying header attachment, units/reference ranges, and comment separation afterward), and

opendataloader-pdf -f markdown-with-html --image-output off

for slide decks (checking for repeated footers, page numbers, and diagram pseudo-tables afterward), with --table-method cluster added only if a structure check shows real improvement. Scanned PDFs are OCR'd first, then converted with opendataloader-pdf. Output is the converted file plus a short report of which route was chosen, whether a retry was needed, whether cleanup or repair was applied, the recommended final output file, and any remaining fidelity limitations - the original PDF itself is always left untouched.

Built around the opendataloader-pdf CLI as the default conversion engine, with pdfinfo and pdftotext -layout used for fast source classification and OCR tooling used as a preprocessing step for scanned or image-only PDFs.

Anyone converting PDFs - medical/lab reports, slide decks, table-heavy financial or business documents, academic papers, or mixed-layout files - into Markdown, HTML, text, JSON, or DOCX who needs the output to preserve real structure (correct label-to-value attachment, intact tables, clean sections) rather than accepting whatever a single default pipeline produces on the first pass.