PDF Text Extractor Skill for Codex

PDF Text Extractor

Optionally collect full-text snippets to deepen evidence beyond abstracts.

This skill is intentionally conservative: in many survey runs, abstract/snippet mode is enough and avoids heavy downloads.

Inputs

• papers/core_set.csv (expects paper_id, title, and ideally pdf_url/arxiv_id/url)
• Optional: outline/mapping.tsv (to prioritize mapped papers)

Outputs

• papers/fulltext_index.jsonl (one record per attempted paper)
• Side artifacts:
• papers/pdfs/<paper_id>.pdf (cached downloads)
• papers/fulltext/<paper_id>.txt (extracted text)

Decision: evidence mode

• queries.md can set evidence_mode: "abstract" | "fulltext".
• abstract (default template): do not download; write an index that clearly records skipping.
• fulltext: download PDFs (when possible) and extract text to papers/fulltext/.

Local PDFs Mode

When you cannot/should not download PDFs (restricted network, rate limits, no permission), provide PDFs manually and run in “local PDFs only” mode.

• PDF naming convention: papers/pdfs/<paper_id>.pdf where <paper_id> matches papers/core_set.csv.
• Set - evidence_mode: "fulltext" in queries.md.
• Run: uv run python .codex/skills/pdf-text-extractor/scripts/run.py --workspace <workspace> --local-pdfs-only

If PDFs are missing, the script writes a to-do list:

• output/MISSING_PDFS.md (human-readable summary)
• papers/missing_pdfs.csv (machine-readable list)

Workflow (heuristic)

• Read papers/core_set.csv.
• If outline/mapping.tsv exists, prioritize mapped papers first.
• For each selected paper (fulltext mode):
• resolve pdf_url (use pdf_url, else derive from arxiv_id/url when possible)
• download to papers/pdfs/<paper_id>.pdf if missing
• extract a reasonable prefix of text to papers/fulltext/<paper_id>.txt
• append/update a JSONL record in papers/fulltext_index.jsonl with status + stats
• Never overwrite existing extracted text unless explicitly requested (delete the .txt to re-extract).

Quality checklist

• [ ] papers/fulltext_index.jsonl exists and is non-empty.
• [ ] If evidence_mode: "fulltext": at least a small but non-trivial subset has extracted text (strict mode blocks if extraction coverage is near-zero).
• [ ] If evidence_mode: "abstract": the index covers every papers/core_set.csv paper and every record clearly reflects skip_mode_abstract (no downloads attempted). fulltext_max_papers does not truncate this zero-download index.

Script

Quick Start

• uv run python .codex/skills/pdf-text-extractor/scripts/run.py --help
• uv run python .codex/skills/pdf-text-extractor/scripts/run.py --workspace <workspace>

All Options

• --max-papers <n>: cap number of papers processed (can be overridden by queries.md)
• --max-pages <n>: extract at most N pages per PDF
• --min-chars <n>: minimum extracted chars to count as OK
• --sleep <sec>: delay between downloads
• --local-pdfs-only: do not download; only use papers/pdfs/<paper_id>.pdf if present
• queries.md supports: evidence_mode, fulltext_max_papers, fulltext_max_pages, fulltext_min_chars

Examples

• Abstract mode (no downloads):
• Set - evidence_mode: "abstract" in queries.md, then run the script (it will emit papers/fulltext_index.jsonl with skip statuses)
• Fulltext mode with local PDFs only:
• Set - evidence_mode: "fulltext" in queries.md, put PDFs under papers/pdfs/, then run: uv run python .codex/skills/pdf-text-extractor/scripts/run.py --workspace <workspace> --local-pdfs-only
• Fulltext mode with smaller budget:
• uv run python .codex/skills/pdf-text-extractor/scripts/run.py --workspace <workspace> --max-papers 20 --max-pages 4 --min-chars 1200

Notes

• Downloads are cached under papers/pdfs/; extracted text is cached under papers/fulltext/.
• The script does not overwrite existing extracted text unless you delete the .txt file.

Troubleshooting

Issue: no PDFs are available to download

Fix:

• Use evidence_mode: abstract (default) or provide local PDFs under papers/pdfs/ and rerun with --local-pdfs-only.

Issue: extracted text is empty/garbled

Fix:

• Try a different extraction backend if supported; otherwise mark the paper as abstract evidence level and avoid strong fulltext claims.