PDF Skill - PDF to Markdown Conversion

Scheduling

Goal

Convert PDF files into structured Markdown or another requested extraction format while preserving readable document structure for LLM context, RAG, or downstream review.

Intent signature

• User asks to convert, parse, read, extract, or transform a PDF.
• User needs PDF text, headings, lists, tables, or images prepared for AI consumption.
• User mentions "PDF to markdown", "parse PDF", "read this PDF", or equivalent wording.

When to use

• Converting PDF documents to Markdown for LLM context or RAG
• Extracting structured content such as tables, headings, lists, images, footnotes, or hyperlinks
• Preparing PDF data for AI consumption
• Checking whether a PDF has a text layer before choosing OCR

When NOT to use

• Generating or creating PDFs -> use document-generation tools
• Editing existing PDFs -> out of scope
• Reading an already-text file -> use direct file reading
• Processing HWP, HWPX, DOCX, XLSX, or slide decks -> use the matching document skill

Expected inputs

• input_path: PDF file or folder path
• output_dir: optional target directory
• format: optional output format, default markdown
• ocr_languages: optional OCR language list for scanned or image-based PDFs
• extraction_options: optional flags for tagged structure, image extraction, or hybrid conversion

Expected outputs

• Markdown, text, JSON, HTML, or combined extraction output
• Normalized Markdown when Markdown is produced
• A short report with output path, page count, and conversion issues

Dependencies

• uvx opendataloader-pdf for standard conversion
• uvx --from "opendataloader-pdf[hybrid]" opendataloader-pdf-hybrid for OCR or hybrid conversion (the hybrid server is a console script of the [hybrid] extra, not a standalone package)
• uvx mdformat for Markdown normalization
• Local filesystem access to input and output paths
• Optional OCR runtime via the hybrid server

Control-flow features

• Branches on text-layer quality, tagged PDF availability, scan/OCR needs, and user-requested output format
• Calls external CLI tools through uvx
• Reads local files and writes local extraction outputs
• Uses a hybrid server only when OCR or complex extraction needs justify it

Structural Flow

Entry

1. Confirm that the input path exists and is a PDF file, PDF folder, or supported batch input.
2. Check file size and warn when the input is large enough to risk slow conversion or memory pressure.
3. Resolve output_dir and the expected output filename.

Scenes

1. PREPARE: Validate the input path, output target, and requested extraction options.
2. ACQUIRE: Assess whether the PDF has a readable text layer by extracting a text preview.
3. ACT: Convert using standard mode, tagged-structure mode, or hybrid OCR mode.
4. VERIFY: Run mdformat for Markdown output and inspect the result for readable structure.
5. FINALIZE: Report output path, page count, format, and any extraction quality issues.

Transitions

• If the preview text is readable, use standard conversion.
• If the PDF is tagged and standard output is garbled, retry with --use-struct-tree.
• If tables are missing or broken, retry with --table-method cluster or --markdown-with-html before escalating to hybrid mode.
• If the PDF is scanned or image-based, start or reuse the hybrid OCR server and convert with hybrid mode.
• If conversion fails because the PDF is encrypted, stop and ask for the password or an unlocked copy.
• If conversion hits memory or size limits, process smaller page ranges into distinct output directories (or append --to-stdout) so repeated runs do not overwrite the same basename.

Failure and recovery

Failure	Recovery
uvx unavailable	Ask user to install uv before conversion
opendataloader-pdf-hybrid not found	Invoke via uvx --from "opendataloader-pdf[hybrid]" opendataloader-pdf-hybrid; the bare package name does not exist on PyPI
Password-protected PDF	Ask for password or unlocked PDF
Garbled output	Retry with tagged structure or hybrid mode
Missing tables	Retry with --table-method cluster or --markdown-with-html first; hybrid mode for scanned tables
OCR language mismatch	Retry with explicit OCR languages, for example ko,en
Large file or memory pressure	Split into page ranges using distinct output directories or --to-stdout; never reuse one output directory for the same basename

Exit

• Success: output file exists, Markdown is formatted when applicable, and extracted structure is readable.
• Partial success: output exists but quality issues are reported explicitly.
• Failure: no reliable output is produced and the blocking cause is reported.

Logical Operations

Actions

Action	SSL primitive	Evidence
Validate path and options	VALIDATE	Input preflight in execution protocol
Probe text layer	READ	Text preview extraction
Choose conversion strategy	SELECT	Standard, tagged, or hybrid mode decision
Run converter	CALL_TOOL	uvx opendataloader-pdf
Start OCR server	CALL_TOOL	uvx --from "opendataloader-pdf[hybrid]" opendataloader-pdf-hybrid
Write output artifact	WRITE	Markdown, text, JSON, or HTML output
Normalize Markdown	CALL_TOOL	uvx mdformat
Inspect extraction quality	VALIDATE	Structure/readability verification
Report result	NOTIFY	Final user-facing summary

Tools and instruments

• opendataloader-pdf: primary PDF extraction CLI
• opendataloader-pdf-hybrid: hybrid OCR and complex extraction path
• mdformat: Markdown normalization
• Filesystem commands such as file, wc, or pdfinfo may be used for preflight when available

Canonical command path

uvx opendataloader-pdf "{input_path}" -f text --pages 1-3 --to-stdout -q   # text-layer probe (no temp files)
uvx opendataloader-pdf "{input_path}" --format markdown --output-dir "{output_dir}"
uvx mdformat "{output_path}"

For scanned/image-based PDFs, start OCR first and then convert through hybrid mode:

uvx --from "opendataloader-pdf[hybrid]" opendataloader-pdf-hybrid --port 5002 --force-ocr --ocr-lang "{languages}"
uvx opendataloader-pdf --hybrid docling-fast --hybrid-mode full "{input_path}" --format markdown --output-dir "{output_dir}"

On-request flags — --sanitize (PII masking), --pages (range extraction), --threads (parallel pages), --detect-strikethrough, hybrid --enrich-formula / --enrich-picture-description — are detailed in resources/execution-protocol.md. The two enrichment flags require client-side --hybrid-mode full.

Resource scope

Scope	Resource target
LOCAL_FS	Input PDFs and generated output files
PROCESS	uvx subprocesses and optional hybrid server
MEMORY	Extracted previews and validation notes
OTHER	OCR model/runtime behavior inside hybrid mode

Preconditions

• The input PDF path exists and is readable.
• The output location is writable or can be created.
• Required CLIs are available through uvx.
• OCR is only attempted when hybrid mode is available or can be started.
• Defaults come from the pdf: section of .agents/oma-config.yaml, falling back to config/pdf-config.yaml for unset keys; explicit user options override both, and output.overwrite: false requires confirmation before replacing an existing output file.

Effects and side effects

• Creates or overwrites extraction output depending on configuration and user intent.
• May start a local hybrid OCR server on the configured port.
• May consume significant CPU, memory, or time for large or scanned PDFs.
• Does not intentionally modify the source PDF.

Guardrails

1. Do not invent missing content when extraction is incomplete.
2. Always report garbled text, missing tables, OCR uncertainty, or partial extraction.
3. Prefer standard conversion first when the text layer is readable.
4. Use OCR only when the PDF is scanned, image-based, or standard extraction quality is insufficient.
5. Keep detailed command sequences in resources/execution-protocol.md rather than duplicating every variant here.

References

• Execution protocol: resources/execution-protocol.md
• Configuration: read the pdf: section of .agents/oma-config.yaml first, then fall back to config/pdf-config.yaml for any key it does not set. User overrides belong in .agents/oma-config.yaml, since oma update overwrites the skill config.
• Context loading: ../_shared/core/context-loading.md
• Quality principles: ../_shared/core/quality-principles.md