Route PDF conversions through analysis-driven extraction

PDF Conversion Router picks the extraction tools, CLI flags, and cleanup strategy for converting a PDF into another format, instead of running one fixed default pipeline. It classifies the source PDF (native digital text, noisy OCR, image-only/scanned, slide deck, medical or lab report, table-heavy business document, narrative article, or mixed layout), classifies the target output (markdown-with-html, markdown, html, text, json, or docx), and chooses the strongest route for that combination before validating the result on representative sections and retrying with better settings if needed. opendataloader-pdf is the default primary conversion engine for every task; other tools are used only to classify the PDF, preprocess with OCR, validate against layout-preserving text, or repair noisy output - never as the default replacement route. When output still needs a pass after conversion, the fix is sorted into one of three buckets: cleanup (removing noise like repeated footers or stray symbols without changing meaning), structural correction (repairing attachment when the extractor found the right content but the wrong shape, such as a flattened table or notes merged into result rows), or route retry (the extraction path itself was wrong, so no amount of cleanup will fix it) - the least invasive bucket that produces a faithful result is preferred, but raw noisy output is never left untouched if it is clearly improvable.

Use it whenever a PDF needs converting to .md, .html, .txt, .json, .docx, or structured notes, especially when the PDF is scanned, OCR-heavy, table-heavy, slide-based, medical, academic, or multi-column. It explicitly does not promote one flag combination into a universal default because it worked on a single PDF - document-specific evidence outranks habit, and a heuristic like --table-method cluster is a targeted default for medical PDFs, not an automatic choice for slide decks or every business document. Throughout, document fidelity - correct content, correct attachment between labels and values, correct section structure - is optimized ahead of visual fidelity; a structurally correct, readable result beats one that looks like the original page but is semantically broken.

Input is the source PDF and the desired output format. For a medical or lab report, the default route is:

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

followed by validation that table headers, and the attachment of value, unit, and reference range, survived intact. Output is the converted file, written to a dedicated output folder while the original PDF is left untouched, plus a delivery summary that states the final chosen output path (not just a command summary): which route was chosen, whether a retry was needed, whether a cleanup or structural-repair pass was applied, the recommended final output file, and any remaining fidelity limitations. Validation always checks one early section and one structurally difficult section, never just page 1 or the top of the file, and it verifies the actual rendered structure rather than trusting a successful command exit code - red flags include flattened table rows, labels detached from values, units drifting into adjacent rows, and a good exit code paired with visibly poor structure.

opendataloader-pdf CLI as the primary engine, with pdfinfo and pdftotext -layout for quick classification and validation, and an OCR pass for image-only or scanned PDFs before conversion. Retries follow a fixed order: same engine with better flags, then a different output shape, then hybrid/full mode, then a cleanup pass, then OCR plus reconversion, and only as a last resort a non-OpenDataLoader fallback.

Anyone converting PDFs where structure matters as much as text - medical or lab reports where values must stay attached to the right exam name, table-heavy financial documents, slide decks with pseudo-tables and diagrams, or any pipeline that has been burned by a PDF converter that reports success while silently flattening or fusing the content that mattered.