Fetching Papers

Fetching Papers

Overview

Download an academic paper and parse it to markdown + extracted figures using mineru, storing results under ./papers/<PAPER-ID>/ with the project's existing convention.

Core principle: Paper-ID matches the project's code-dir name when one exists (e.g., DiffV2IR/ ↔ papers/DiffV2IR/). Never invent a new naming scheme if the project already has one.

When to Use

• User says "download the X paper", "get the paper for Y", "parse this PDF"
• User references a paper by name/abbreviation that isn't yet in ./papers/
• User wants markdown output of a PDF to read or feed to another tool

Skip when: the paper already exists at papers/<ID>/<ID>.md — just read it.

Workflow

digraph fetch {
  "Identify paper" [shape=box];
  "Resolve to canonical URL" [shape=box];
  "papers/<ID>/ exists?" [shape=diamond];
  "Skip — read existing" [shape=box];
  "Download PDF" [shape=box];
  "Parse with mineru" [shape=box];
  "Verify output" [shape=box];

  "Identify paper" -> "Resolve to canonical URL";
  "Resolve to canonical URL" -> "papers/<ID>/ exists?";
  "papers/<ID>/ exists?" -> "Skip — read existing" [label="yes, .md present"];
  "papers/<ID>/ exists?" -> "Download PDF" [label="no"];
  "Download PDF" -> "Parse with mineru";
  "Parse with mineru" -> "Verify output";
  "Verify output" -> "Update .gitignore";
  "Update .gitignore" -> "Append to PAPERS.md";
}

Step 1 — Identify the paper & pick Paper-ID

The user often names a paper by abbreviation (e.g., "PID", "F-ViTA", "DiffV2IR"). Resolve it:

1. Check existing project layout first — ls papers/ and ls the project root.
   • If a top-level code dir matches (e.g., PID/), use that name as Paper-ID (preserve case).
   • Look at sibling papers/<other>/ entries for the naming style already used.
2. Grep for citations in the matching code dir's README, paper-related comments, or *.bib:

   Grep pattern="arxiv|doi|openreview|huggingface.co/papers" path="<code-dir>"
   

3. If no local reference, ask the user for the arXiv ID / URL / title. Do not guess IDs.

Step 2 — Resolve to a canonical PDF URL

Try sources in order (use the first that works):

If the user has the huggingface-skills:huggingface-papers skill available, prefer it for HF-indexed papers (gives structured metadata + abstract).

Step 3 — Ensure mineru is installed

Check the project's Python env first:

• E:\Code\CMU\lab\RGB2IR stores env details in python_environment.md — read it and use that env.
• Otherwise look for pyproject.toml / .venv/ / requirements.txt.
• If the user hasn't specified an env, ask before installing.

Install via uv into that env:

# Preferred: install into the project/user-preferred env
uv tool install -U "mineru[core]"
# OR, for one-off use without polluting tool env:
uv run --with "mineru[core]" -- mineru --help

Verify: mineru --version. If it fails, do not silently fall back to another parser — report the error and ask.

Step 4 — Download the PDF

mkdir -p "papers/<ID>"
curl -L -f -o "papers/<ID>/<ID>.pdf" "<resolved-url>"

Use -f so HTTP errors fail the command. Verify file is a real PDF: file papers/<ID>/<ID>.pdf should report PDF document.

Step 5 — Parse with mineru

mineru -p "papers/<ID>/<ID>.pdf" -o "papers/<ID>/"

mineru writes markdown + figures into papers/<ID>/<ID>/ (nested subdir named after the PDF). That nested layout matches what already exists for papers/F-ViTA/F-ViTA/.

Common flags:

• -b pipeline (default, CPU/GPU local) vs -b vlm-transformers (heavier, better for complex layout)
• -l en to pin language and avoid autodetection overhead
• --formula-enable true --table-enable true (defaults on; keep on for ML papers)

Run mineru --help for the full list.

Step 6 — Verify

• Read the first ~50 lines of the generated .md and confirm title + authors match what the user asked for.
• List the output dir so the user sees what was produced.
• Never claim success without opening the markdown.

Step 7 — Manage .gitignore (if inside a git repo)

mineru outputs (PDFs, extracted figures, layout debug PDFs, JSON content dumps) are large and binary. They must not be committed.

1. Detect git: check for a .git/ directory at or above the project root (e.g., git rev-parse --show-toplevel). If not a git repo, skip this step.

2. Ensure .gitignore at the repo root contains the following block — append it if missing (do not duplicate existing lines):

   # Papers: large binaries & mineru outputs — tracked via papers/PAPERS.md
   papers/**/*.pdf
   papers/**/images/
   papers/**/*_content_list.json
   papers/**/*_layout.pdf
   papers/**/*_middle.json
   papers/**/*_model.json
   papers/**/*_origin.pdf
   papers/**/*_spans.pdf
   

   Keep papers/**/*.md tracked — the parsed markdown is small and useful in-repo.

3. Verify: git check-ignore -v papers/<ID>/<ID>.pdf should report the rule matched.

Step 8 — Update papers/PAPERS.md

Because PDFs and figures are gitignored, papers/PAPERS.md is the single source of truth recording what's been fetched, where it came from, and how to re-fetch it.

1. If papers/PAPERS.md does not exist, create it using the template below.
2. Append a new entry for this paper. If an entry with the same Paper-ID already exists, update it rather than duplicating.
3. Keep entries sorted alphabetically by Paper-ID.

PAPERS.md template

# Papers Index

This directory stores parsed academic papers. Large files (PDFs, figure images, mineru intermediates) are gitignored — only the parsed `.md` is tracked. Use this index to know which papers have been fetched and how to re-download them.

Re-fetch any paper with the `fetching-papers` skill using its source URL below.

---

## <Paper-ID>

- **Title:** <full paper title>
- **Authors:** <Author1, Author2, et al.>
- **Year / Venue:** <e.g., 2024 / CVPR>
- **Source:** <arXiv | HuggingFace | OpenReview | CVF | ACL | DOI>
- **Source URL:** <canonical landing page, e.g., https://arxiv.org/abs/2401.12345>
- **PDF URL:** <direct PDF link used to download>
- **arXiv ID / DOI:** <id, if applicable>
- **Local path:** `papers/<Paper-ID>/<Paper-ID>/<Paper-ID>.md`
- **Related code dir:** `<./code-dir/>` or `—`
- **Fetched:** <YYYY-MM-DD>
- **Parser:** mineru <version> (backend: `pipeline` | `vlm-transformers`)
- **Summary:** <1–3 sentence TL;DR — what the paper proposes and why this project cares>
- **Tags:** `<tag1>`, `<tag2>`  *(e.g., `rgb2ir`, `diffusion`, `dataset`)*

---

Example filled entry

## DiffV2IR

- **Title:** DiffV2IR: Visible-to-Infrared Image Translation via Diffusion Models
- **Authors:** Jane Doe, John Smith, et al.
- **Year / Venue:** 2024 / arXiv preprint
- **Source:** arXiv
- **Source URL:** https://arxiv.org/abs/2403.12345
- **PDF URL:** https://arxiv.org/pdf/2403.12345.pdf
- **arXiv ID / DOI:** 2403.12345
- **Local path:** `papers/DiffV2IR/DiffV2IR/DiffV2IR.md`
- **Related code dir:** `./DiffV2IR/`
- **Fetched:** 2026-04-22
- **Parser:** mineru 1.x (backend: pipeline)
- **Summary:** Proposes a diffusion-based pipeline for translating visible-spectrum images to infrared, conditioning on structural edges to preserve thermal-consistent geometry.
- **Tags:** `rgb2ir`, `diffusion`, `image-translation`

Final layout

papers/
  PAPERS.md              ← tracked; index of all fetched papers
  <ID>/
    <ID>.pdf             ← gitignored
    <ID>/
      <ID>.md            ← tracked
      images/            ← gitignored
      <ID>_content_list.json  ← gitignored
      <ID>_layout.pdf    ← gitignored (optional debug)

Common Mistakes

Red Flags — STOP

• About to pip install instead of uv tool install
• About to save as flat papers/foo.pdf
• About to guess an arXiv ID without asking
• About to fall back to a non-mineru parser silently
• About to create a new venv without consulting python_environment.md
• About to git add a PDF or a papers/<ID>/<ID>/images/ dir
• About to mark the task complete without updating papers/PAPERS.md

All of these mean: stop, re-read Steps 1 and 3.