bibvet

Vet your bibliography. Catches hallucinated citations, wrong years, and fabricated DOIs.

bibvet reads .bib files and verifies every entry against CrossRef, Semantic Scholar, and arXiv. It cross-checks the DOI a paper claims against the title it claims to have — catching the most common LLM citation hallucinations — and can write a cleaned .bib with canonical fields.

Install

# uv (recommended) — installs the `bibvet` command
uv tool install git+https://github.com/dipta007/bibvet.git

# or run without installing
uvx --from git+https://github.com/dipta007/bibvet.git bibvet refs.bib

# or pipx
pipx install git+https://github.com/dipta007/bibvet.git

# or plain pip
pip install git+https://github.com/dipta007/bibvet.git

No API keys, no config, no setup needed.

Quick start

bibvet refs.bib                # check one file
bibvet refs.bib --strict       # high-recall sweep (recommended pre-submission)
bibvet refs.bib --fix          # also write refs.fixed.bib with canonical fields

What it catches

• Hallucinated DOIs — fabricated doi: strings that don't exist
• Wrong year — paper exists, but the cited year is off
• Wrong / swapped authors — first author is missing, or author order is wrong
• Mismatched title and DOI — DOI resolves to one paper, title matches another (the LLM mixed two real papers)
• Fully hallucinated entries — paper doesn't exist at any source
• Venue mismatches — softer signal, often cosmetic
• Cosmetic differences — Y. Bengio vs Yoshua Bengio (informational only)

Usage

bibvet refs.bib                        # check one file
bibvet refs1.bib refs2.bib refs3.bib   # multiple files
bibvet ./papers/                       # all .bib files in a directory
bibvet ./papers/ --recursive           # also search subdirectories
bibvet refs.bib --fix                  # additionally write refs.fixed.bib
bibvet refs.bib --explain mycite2024   # detail for one entry
bibvet refs.bib --format md            # write report.md (clickable links)
bibvet refs.bib --format json          # machine-readable
bibvet refs.bib --skip-source arxiv    # disable a source if it's rate-limiting
cat refs.bib | bibvet -                # read from stdin

Exit codes

• 0 — everything verified or skipped
• 1 — warnings only
• 2 — at least one entry needs attention

CI-friendly: drop bibvet refs.bib into a pre-submission check and fail the build on 2.

Output

refs.bib · 23 entries · 20 verified · 2 fixable · 1 warning
  bad-cite (refs.bib:42)
     https://doi.org/10.5555/3295222.3295349
     venue: Advances in Neural Information Processing Systems (2017)
     error doi: '10.1234/fake' -> '10.5555/3295222.3295349' (DOI does not match canonical record)
     error year: '2018' -> '2017' (year mismatch)

--fix writes <name>.fixed.bib next to each input with only the fields bibvet verifies: author, title, year, journal/booktitle, doi, eprint. Unverified fields like pages, volume, number are dropped — they're commonly hallucinated and we can't confirm them. Entries that can't be auto-fixed (cross-check failures, unverified) are left byte-identical with a % bibvet: comment above them — grep '^% bibvet:' refs.fixed.bib to find what needs review.

Strict mode

bibvet refs.bib --strict

For high-recall hallucination sweeps. Promotes warnings to errors:

• Any title difference flagged (was: cosmetic differences silenced)
• Venue mismatches treated as errors (was: warnings)
• Books, theses, and @misc entries with no canonical match flagged (was: silently skipped)
• Single-source matches noted for manual review

Slightly more false positives, but won't miss subtly altered citations.

Configuration

bibvet works with no config. Optional environment variables:

• SEMANTIC_SCHOLAR_API_KEY — higher quota for Semantic Scholar (1 req / 50ms vs. 1 req / second without)
• CROSSREF_MAILTO — opts into CrossRef's polite pool

Getting a Semantic Scholar API key

If you're verifying more than a few dozen entries, the unauthenticated rate limit (~100 requests / 5 min) will throttle you. The key is free.

1. Go to https://www.semanticscholar.org/product/api#api-key-form and request a key.

2. Approval typically takes a few business days; they email you the key.

3. Set it in your shell:

   export SEMANTIC_SCHOLAR_API_KEY=sk-...

   Or persist it in ~/.zshrc / ~/.bashrc. bibvet picks it up automatically — no flags needed.

Cache lives at the platform-default user cache directory (~/Library/Caches/bibvet/ on macOS, ~/.cache/bibvet/ on Linux). 30-day TTL. bibvet cache clear to purge.

License

MIT.