resilient-pdf（resilient PDF 搜索）

Resilient PDF

Use this skill as a fallback workflow for PDFs that break normal analysis paths.

Overview

Prefer the built-in pdf tool first when it is likely to work. If it fails, hangs, times out, or the file is too large, switch to this local workflow.

Read references/patterns.md if you need the rationale, chunking heuristics, or fallback guidance.

Workflow

1. Confirm the PDF source.

- If remote, download it into the workspace first. - If local, confirm the path and file size.

1. Decide whether the normal path is already broken.

- Trigger this skill when the built-in pdf tool aborts, provider-native upload fails, or file limits make direct analysis unlikely to work.

1. Run the helper extractor.

- Use scripts/extract_pdf.py to extract markdown locally. - Use --url to download a remote PDF first. - Add --chunk-dir when the output will be too large to read in one pass. - Add --summary-out to generate a lightweight first-pass summary artifact.

1. Inspect the extracted output.

- Read the head, table of contents, or key sections first. - Do not trust a summary until the extracted text looks sane.

1. Summarize or analyze.

- For short outputs, read the extracted markdown directly. - For long outputs, read selected chunks or key sections. - Use the generated first-pass summary as a navigation aid, not as final truth. - Keep quoted claims and numeric claims grounded in the extracted text.

Helper script

Local file command:

python3 skills/resilient-pdf/scripts/extract_pdf.py <file.pdf> --out <output.md> --json

Remote URL command:

python3 skills/resilient-pdf/scripts/extract_pdf.py \
  --url <https://example.com/file.pdf> \
  --out <output.md> \
  --download-to <downloaded.pdf> \
  --json

Chunked plus summary command:

python3 skills/resilient-pdf/scripts/extract_pdf.py <file.pdf> \
  --out <output.md> \
  --chunk-dir <chunk-dir> \
  --summary-out <summary.md> \
  --chunk-chars 120000 \
  --chunk-overlap 4000 \
  --json

The script:

• accepts either a local file path or --url
• downloads remote PDFs when needed
• looks for uvx
• invokes uvx --from 'markitdown[pdf]' markitdown
• writes extracted markdown
• optionally writes chunk files
• optionally writes a lightweight first-pass summary markdown file
• emits a machine-readable JSON result

If dependencies are missing

If uvx is not available, tell the operator the exact command to install it:

python3 -m pip install --user --break-system-packages uv

Do not silently install dependencies unless the user asked you to.

Output expectations

A successful run should give you:

• downloaded PDF path when using --url
• extracted markdown path
• byte count
• text character count
• optional chunk paths
• optional first-pass summary path

Use those outputs as the source of truth for later summarization.

Notes

• This skill does not replace the built-in pdf tool. It is the fallback when that path is unreliable.
• Prefer workspace-local outputs so later reads and summaries are reproducible.
• If the extracted markdown is noisy, inspect section headers and sample passages before making strong claims.