The original visual source. It remains the authority for what appears on each page.
Level 1 · Overview
Last modified: (dates/navigation repair). Content current as of: (last retained content commit; content verification date unavailable). Original content last edited: . Current claims have not been freshly reverified.
How PDFMD converts a PDF into structured Markdown
PDFMD reconstructs the logical document—text, paragraphs, lists, notes, tables, page relationships, and meaningful visual evidence—without trying to reproduce the page design.
Original PDFpage description
→
Read evidencetext, pixels, geometry
→
Record structurePDFMD
→
RenderMarkdown
→
Readerclean or annotated view
The deployed converter performs this route with deterministic orchestration and OCR. It does not call an LLM.
The problem
A PDF describes pages, not necessarily a document
A PDF reliably records marks and positions. It may not state that two lines are one paragraph, a number is a page label, or a note belongs to a marker in the body. Recognition infers those relationships and records the source evidence for each decision.
Illustrative example. A numbered paragraph begins on page 2, contains a two-item list and a bold phrase, passes a footer, and ends on page 3 with a footnote call. Later pages use the same example to explain each method. It is explanatory, not verified recognizer output.
The records
PDF, PDFMD, and Markdown
PDFMD
A document record written in extended Markdown. It stores meaning, page provenance, uncertainty, and asset references.
Markdown
A disposable reading view generated from PDFMD. Changing view settings does not change the canonical record.
Recognition bundle
Detailed observations used for diagnosis and evaluation. The current browser download contains PDFMD, human Markdown, and referenced assets; development runs may retain a separate bundle.
Responsibility
Script, model, and human roles
ActorStateRole
Code and OCRImplementedInput checks, capture, OCR, detectors, validation, packaging, and rendering.
Language model requestPlannedA bounded second opinion for one ambiguous decision. Not a deployed dependency.
Human reviewerWhen requiredConstructs references, resolves high-impact ambiguity, and accepts or rejects releases. The converter does not currently pause for an interactive decision.
Documentation map
Continue by subject
Historical recognition gate
Why English text conservation is blocked, what the evidence shows, and what must happen next.
Conversion
Input checks, native and scanned routes, evidence, semantic decisions, PDFMD, and output.
Recognition methods
Paragraphs, lists, type, notes, tables, page furniture, and graphics.
Review and decisions
Supported assertions, unresolved alternatives, bounded model requests, and human corrections.
PDFMD and rendering
The canonical format and the generation of different Markdown views.
Operations
Live conversion, background analysis, retention, and deployment boundaries.
Development and evaluation
Human-built references, two freezes, correction cycles, and independent evaluation.