ARI Docs

English

日本語
中文

English

日本語
中文

Appearance

Importing external papers ​

The paper registry (~/.ari/paper_registry/) holds external papers the PaperBench wizard can run audits against. This page covers the four import paths and license handling.

Storage layout ​

{ARI_PAPER_REGISTRY_DIR or ~/.ari/paper_registry}/
├── manifest.jsonl            # one paper per line (JSON)
└── papers/
    └── <paper_id>/
        ├── paper.pdf         # source PDF (optional but recommended)
        ├── ad.pdf            # artifact description (optional)
        └── ae.pdf            # artifact evaluation (optional)

Override the root with ARI_PAPER_REGISTRY_DIR.

Import paths ​

arXiv ID ​

Most-used path. Fill the wizard with source_type=arxiv and source=2404.14193. Auto-fetch of metadata + PDF from arXiv is deferred to a follow-up release; for v0.7.2 you still need to type title / authors / license by hand.

bash
curl -X POST http://localhost:8765/api/paperbench/papers/import \
  -H 'Content-Type: application/json' \
  -d '{
    "source_type": "arxiv",
    "source": "2404.14193",
    "title": "LLAMP: assessing latency tolerance",
    "license": "CC BY 4.0",
    "authors": ["Alice", "Bob"],
    "year": 2024,
    "venue": "SC24",
    "artifact_url": "https://github.com/spcl/llamp"
  }'

DOI ​

Same form as arXiv but source_type=doi, source=10.1109/<conf>.YYYY.NNNNN (e.g. an ACM DL or IEEE Xplore DOI for an SC, OSDI, or USENIX paper). Use for IEEE/ACM papers that are not on arXiv.

Upload (local PDF) ​

source_type=upload. Use /api/upload to stash the PDF first, then pass its path as pdf_path:

bash
TMP=$(curl -F 'file=@./mypaper.pdf' http://localhost:8765/api/upload | jq -r .path)
curl -X POST http://localhost:8765/api/paperbench/papers/import \
  -H 'Content-Type: application/json' \
  -d "{
    \"source_type\": \"upload\",
    \"source\": \"local-upload-$(date +%s)\",
    \"title\": \"My SC24 camera-ready\",
    \"license\": \"IEEE Author proprietary\",
    \"pdf_path\": \"$TMP\"
  }"

The PDF is copied into papers/<paper_id>/paper.pdf. AD / AE appendices follow the same path under ad_pdf_path / ae_pdf_path.

Local path ​

source_type=local. For PDFs already on disk where you don't need ARI to re-host them — the registry entry just points at the existing path.

License classification ​

License strings are normalized (lower-cased, whitespace-stripped) and classified into a 4-quadrant assessment:

StatusExamples
usable (permissive AND redistributable)MIT, Apache-2.0, BSD-2/3-Clause, CC0, CC BY, CC BY-SA, CC BY-NC, arXiv non-exclusive
permissive only (NOT redistributable)(none currently — kept as a placeholder)
NOT usableproprietary, IEEE author, ACM author, "all rights reserved", unknown strings

The classification is heuristic and advisory. Final legal review remains the user's responsibility. The GUI shows a green ✅ badge for usable, amber ⚠ for not — both still let the registration go through.

Inspect a paper's assessment:

bash
curl http://localhost:8765/api/paperbench/papers/<paper_id>/license

Duplicate detection ​

Imports with the same paper_id (default: sanitized source) are blocked unless overwrite=true is passed:

bash
curl -X POST http://localhost:8765/api/paperbench/papers/import \
  -H 'Content-Type: application/json' \
  -d '{
    "source_type": "arxiv", "source": "2404.14193",
    "title": "LLAMP v2", "license": "CC BY 4.0",
    "overwrite": true
  }'

paper_id is itself sanitized to [A-Za-z0-9._-]{1,64} — any other characters become -.

Deletion ​

bash
curl -X POST http://localhost:8765/api/paperbench/papers/<paper_id>/delete

Idempotent (returns {deleted: false, reason: "not found"} when the id is unknown).

Metadata patches ​

For fixing typos without losing the registry slot:

bash
curl -X POST http://localhost:8765/api/paperbench/papers/<paper_id>/metadata \
  -H 'Content-Type: application/json' \
  -d '{"venue": "SC25", "year": 2025}'

paper_id is immutable.

See also ​