Text-free sculptural book and optical scanner transforming pages into an organized yellow and cyan knowledge network. — Document extraction is reliable only when the method fits the source and the output is checked against the pages.

Extract Text from PDFs and Scans with Hermes

PDFs can contain selectable text, scanned images, complex tables, equations, or a mixture of all four. The bundled PDF and Document Extraction skill uses a decision path: try direct remote extraction when a public URL is available, use PyMuPDF for lightweight local text and page operations, and reserve marker-pdf for high-quality OCR or complex layouts.

Official skill reference: PDF and Document Extraction

Match the extractor to the source

Remote web_extract is efficient for a public PDF URL and can produce readable Markdown without a local install. PyMuPDF is a good local choice for text-native PDFs, metadata, page ranges, searching, splitting, merging, and image extraction. Marker requires more installation and compute, but it is better suited to scans, tables, equations, and layout-sensitive documents. Its LLM-assisted mode should be used only when the data boundary and provider are acceptable.

Test a representative page before processing hundreds of files. Determine whether the apparent blank output is a scan, an encoding problem, or a permissions issue rather than retrying the same extractor blindly.

Extraction checklist

Record source URL or file hash and page count.
Determine whether text is native, scanned, or mixed.
Test headings, paragraphs, tables, equations, and footnotes.
Limit page ranges during early trials.
Keep OCR and LLM-assisted processing inside the approved privacy boundary.
Preserve page references in downstream notes and citations.
Save extracted images only when they are required and lawful to retain.

Common pitfalls

Assuming selectable text means correct reading order. Multi-column pages can interleave.
Running heavyweight OCR on a simple text PDF. It costs time and can introduce errors.
Trusting tables without visual comparison. Merged cells and footnotes often shift.
Losing page provenance. Plain text without page anchors is difficult to cite or audit.
Uploading private scans to a remote extractor. Confirm the processing boundary first.

Verification

Compare the extracted text with sample pages from the beginning, middle, and end. Check names, dates, totals, minus signs, table headers, equations, and footnotes. Confirm the extracted page count and section order, and search for expected phrases. For high-stakes documents, keep the original page image beside the extracted passage during review; OCR confidence is not a substitute for visual verification.