AgentSkillsCN

read-scientific-pdfs-mineru

使用 MinerU 将科学 PDF 无损转换为 Markdown。前提条件是已预先创建名为 “mineru”的 Conda 环境。

SKILL.md
--- frontmatter
name: read-scientific-pdfs-mineru
description: Extract scientific PDFs to lossless Markdown using MinerU. Assumes a pre-existing conda env named "mineru".

read-scientific-pdfs-mineru

Goal: convert PDFs to Markdown as losslessly as possible (no summarization). This skill assumes a conda env named mineru already exists. If it doesn’t, stop and tell the user what is missing (create the env + install MinerU) and do not attempt installs.

PDF placement

Save the source PDF under agent_assets/<project_name>/artifacts/pdf_conversions/ before extraction. Use the same basename throughout the conversion outputs.

Method differences (choose based on PDF type)

  • •txt: Use for digital PDFs with selectable text. Fastest and most faithful for journals/preprints.
  • •ocr: Use for scanned PDFs or image-only PDFs. Slower; OCR errors are possible.
  • •auto: Let MinerU choose. Good when you’re unsure, but txt is usually best for scientific papers.

Quick use

bash
SKILLS_DIR="$(readlink -f .agent/skills)"
conda activate mineru
python "$SKILLS_DIR/read-scientific-pdfs-mineru/scripts/mineru_extract.py" \
  --filepath_pdf agent_assets/<project_name>/artifacts/pdf_conversions/<paper>.pdf \
  --output_dir agent_assets/<project_name>/pdf_conversions \
  --method txt

If .agent/skills is unavailable, set SKILLS_DIR to your central skills checkout (e.g., $CODEX_HOME/skills) and reuse the same command.

Default execution mode: dispatch via sbatch

Use Slurm by default. Follow AGENTS.md for partition/cpu/mem/account values. Create a small dispatch script in agent_assets/<project_name>/code/ and submit it.

Example agent_assets/<project_name>/code/mineru_dispatch.sbatch:

bash
#!/usr/bin/env bash
#SBATCH --account=<account_name_from_AGENTS>
#SBATCH --time=0-02:00:00
#SBATCH --partition=<partition_from_AGENTS>
#SBATCH --gres=gpu:1
#SBATCH -c <cpu_from_AGENTS>
#SBATCH --mem=<mem_from_AGENTS>
#SBATCH --output=agent_assets/<project_name>/artifacts/mineru_%j.out
#SBATCH --error=agent_assets/<project_name>/artifacts/mineru_%j.err

set -euo pipefail

source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate mineru

SKILLS_DIR="$(readlink -f .agent/skills)"
python "$SKILLS_DIR/read-scientific-pdfs-mineru/scripts/mineru_extract.py" \
  --filepath_pdf agent_assets/<project_name>/artifacts/pdf_conversions/<paper>.pdf \
  --output_dir agent_assets/<project_name>/pdf_conversions \
  --method txt

Submit with:

bash
sbatch agent_assets/<project_name>/code/mineru_dispatch.sbatch

Local fallback (no sbatch)

If the PDF is small and you already have an interactive compute node, run the “Quick use” command directly. This avoids waiting for Slurm scheduling.

If the user explicitly asks to use a compute node (interactive)

SSH to the node in AGENTS.md, activate mineru, then run the script:

bash
ssh <compute_node_name>
source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate mineru
SKILLS_DIR="$(readlink -f .agent/skills)"
python "$SKILLS_DIR/read-scientific-pdfs-mineru/scripts/mineru_extract.py" \
  --filepath_pdf agent_assets/<project_name>/artifacts/pdf_conversions/<paper>.pdf \
  --output_dir agent_assets/<project_name>/pdf_conversions \
  --method txt

Script behavior

  • •Refuses to run unless the active conda env is named mineru.
  • •Refuses to run if the mineru CLI is missing from PATH.
  • •Writes results to the provided --output_dir.
  • •Prints the paths of all extracted .md files.
  • •Optional: --preview-lines N to print the first N lines of each .md for quick context loading.

Recommended output layout

Use agent_assets/<project_name>/pdf_conversions/ so outputs are easy to find and reuse.

Output sanity check

After completion, confirm extraction:

bash
find agent_assets/<project_name>/pdf_conversions -type f -name "*.md"

Streamlined error feedback

  • •Check job state: squeue -j <jobid>
  • •Stream logs: tail -f agent_assets/<project_name>/artifacts/mineru_<jobid>.err
  • •If .err is empty, also check .out.

Seamless handoff (read Markdown into context)

After extraction, open the printed .md file and read it into context. If multiple .md files were produced, pick the one matching the PDF basename.

Example:

bash
sed -n '1,200p' agent_assets/<project_name>/pdf_conversions/<basename>/<subdir>/<basename>.md

If something is missing

If conda activate mineru fails or MinerU isn’t on PATH, tell the user what to create/install and stop.

Throughput (rough guidance)

On GPU nodes, digital PDFs tested so far averaged ~20–60 seconds per page (short papers are slower per page). Expect variability by hardware, PDF layout complexity, and chosen method.

Safety

Never upload PDFs to external services without explicit approval.