Onecite

AI-assisted writing, automated literature pipelines, and copy-paste research habits produce ever more reference objects in ever more formats — and ever more chances for wrong, fabricated, or mismatched bibliographic data. Reference managers (Zotero), parsers (AnyStyle, GROBID), format converters (Citation.js), and identifier-to-BibTeX helpers (doi2bib, Manubot) each solve one slice of the problem. OneCite targets the under-served step before references enter a manuscript, systematic review, or manager: an auditable normalization layer that routes strong identifiers (DOI, PMID, arXiv, ISBN, URL, data/software DOIs) to the applicable metadata services, completes available metadata, reports unresolved entries, and produces BibTeX or CSL-JSON.

OneCite is not another reference manager, and process does not auto-accept fuzzy title matches. Strong identifiers are resolved through documented source routes; explicitly labelled thesis/dissertation citations have a separate OpenAIRE/BASE route and can fall back to fields parsed from the input. Ordinary ambiguous text is returned as ranked candidates through onecite suggest for human review rather than silently emitted as source-resolved output. That process/suggest separation, together with machine-readable JSON/NDJSON, exit codes, and deterministic offline checks, makes OneCite a scriptable building block for agents, batch jobs, and reproducible reviews rather than a GUI library. Source resolution does not establish that a work is authentic, unretracted, or correctly described by upstream metadata.

Features

🌐 Data Sources

Privacy and external services

Normal process and suggest runs can make input-dependent outbound requests. For example, DOI resolution sends the DOI to Crossref (and sometimes a fallback registry); suggest sends citation queries to Crossref, Semantic Scholar, and arXiv; arbitrary URL input is fetched from the supplied host; and thesis queries can be sent to OpenAIRE and BASE. Google Scholar is an optional, scraping-based suggest fallback that is off by default and may be blocked or challenged by a CAPTCHA.

OneCite does not provide a privacy-compliance guarantee or a persistent cache of ordinary live responses. Providers, proxies, output files, and logs may retain data. Review and redact confidential input before use. See Privacy and external services for the exact process/suggest routes, transmitted fields, source-health limits, and the offline benchmark/doctor boundary.

Quick Start

Install and try OneCite in a few steps.

1. Installation

The current public PyPI release is 0.1.1. This working tree documents the unreleased 0.2.0 candidate, so install from the checkout when verifying candidate-only behavior:

# Current stable public release
pip install onecite

# Unreleased 0.2.0 candidate, from the repository checkout
python -m pip install -e .

2. Create an Input File

Create a file named references.txt with your mixed-format references:

# references.txt
# Add blank lines between entries to avoid misidentification

10.1038/nature14539

arXiv:1706.03762

ISBN:9780262035613

https://github.com/tensorflow/tensorflow

10.5281/zenodo.3233118

arXiv:2103.00020

Smith, J. (2020). Neural Architecture Search. PhD Thesis. Stanford University.

3. Run OneCite

Execute the command to process your file and generate a clean .bib output.

onecite process references.txt -o results.bib --quiet

4. View Output

Your results.bib file now contains entries of different types.

@article{LeCun2015Deep,
  doi = "10.1038/nature14539",
  title = "Deep learning",
  author = "LeCun, Yann and Bengio, Yoshua and Hinton, Geoffrey",
  journal = "Nature",
  year = 2015,
  volume = 521,
  number = 7553,
  pages = "436-444",
  publisher = "Springer Science and Business Media LLC",
  url = "https://doi.org/10.1038/nature14539",
  type = "journal-article",
  abstract = "Deep learning allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction...",
}
@inproceedings{Vaswani2017Attention,
  arxiv = "1706.03762",
  title = "Attention Is All You Need",
  author = "Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia",
  year = 2017,
  booktitle = "Advances in Neural Information Processing Systems (NeurIPS)",
  url = "https://arxiv.org/abs/1706.03762",
}
# ... and 5 more entries ...

📖 Advanced Usage

onecite process "10.1038/nature14539"
onecite suggest "Attention is all you need, Vaswani et al., NIPS 2017"
echo "10.1038/nature14539" | onecite process -

Use OneCite directly in your Python scripts.

from onecite import process_references

result = process_references(
    input_content="10.1038/nature14539",
    input_type="txt",
    template_name="journal_article_full",
    output_format="bibtex",
)

print('\n\n'.join(result['results']))

OneCite provides a command-line interface with the following commands and options:

onecite process

The main command for processing references through the OneCite pipeline.

Usage:

onecite process <input_file> [OPTIONS]

Arguments:

• input_file - Input file path, - for stdin, or a strong identifier/reference string

Options:

Examples:

# Process a text file
onecite process references.txt -o results.bib

# Process a BibTeX file with auto-detection
onecite process references.bib

# Use stdin
echo "10.1038/nature14539" | onecite process -

# Process a direct string (DOI)
onecite process "10.1038/nature14539"

# Process with custom template
onecite process references.txt --template conference_paper

# Quiet mode for scripts
onecite process references.txt -o results.bib --quiet

# Automation-friendly JSON with unresolved-entry exit-code handling
onecite process references.txt --json --fail-on-unresolved

# Streaming NDJSON for automation
onecite process references.txt --ndjson

# CSL-JSON item output (a development fixture verifies Pandoc 3.10 consumption)
onecite process references.txt --output-format csl-json -o references.json

The development evidence verifies Pandoc 3.10 consumption of representative emitted items. Quarto, standalone citeproc, and reference-manager import workflows are not separately validated in this release.

Report fields. Beyond results and failed_entries, the processing report carries two audit signals:

• warnings — non-blocking review warnings on resolved entries. Most importantly text_metadata_mismatch: the input text around a resolved DOI appears to describe a different work (the classic hallucinated title+DOI pairing). The DOI remains the resolved identifier and the entry resolves, but it is flagged for review instead of silently emitted as clean output.
• duplicates — the same work appeared more than once in the batch (bare DOI, PMID, formatted citation). It is emitted once; repeats are reported with the emitted entry's cite key.

Failed entries carry the original input excerpt (raw_text) and a reason code — doi_not_found (no registry record was returned after the implemented fallback), no_strong_identifier (ambiguous text; use onecite suggest), source_error (a source/identity failure surfaced on that route), and more. These codes make important cases distinguishable, but they are not a complete provider trace: some PMID, ISBN, and DataCite request errors currently collapse into the same unresolved reason as a lookup miss.

onecite suggest

Search for candidate matches without producing BibTeX or returning a validation passed status.

onecite suggest "Attention is all you need, Vaswani et al., NIPS 2017" --json

Candidates are for review, not source-resolved citations. Each suggestion discloses the health of the consulted scholarly indexes in a sources list. If a source was rate-limited or errored, the suggestion status becomes candidates_found_incomplete / no_candidates_incomplete — the correct match may be missing from the list entirely, and the candidate list must not be treated as exhaustive. Candidates whose year contradicts the year cited in the query are penalized and flagged with year_conflict. To turn a reviewed candidate into source-resolved BibTeX, resolve its DOI through onecite process "<doi>".

Optional Google Scholar fallback. suggest accepts --google-scholar (requires the optional scholarly package: pip install onecite[scholar]). It is consulted only as a best-effort fallback when CrossRef and Semantic Scholar return nothing. Because it scrapes a service with no public API, it is off by default, may be rate-limited or blocked by a CAPTCHA, and is not guaranteed to be reproducible — it is exposed only on suggest (candidates for human review), never on process.

pip install onecite[scholar]
onecite suggest "some obscure title" --google-scholar

onecite --version

Display the installed OneCite version.

Usage:

onecite --version

onecite version

Alternative command to display version information.

Usage:

onecite version

onecite templates

List the bundled fallback BibTeX templates and the fields they request.

Usage:

onecite templates
onecite templates --json

onecite benchmark

Run a small deterministic regression suite for covered DOI lookup, arXiv lookup, PMID/PubMed lookup, GitHub software URLs, Zenodo/DataCite dataset DOIs, and mixed valid/invalid batches. The command is designed for CI and automation workflows that need a machine-readable pass/fail check; it is not a comprehensive citation-accuracy benchmark.

Usage:

onecite benchmark [OPTIONS]

Options:

Examples:

onecite benchmark
onecite benchmark --json
onecite benchmark --live --json
onecite benchmark --cases my_cases.json --min-success-rate 1.0 --json
onecite benchmark --anti-hallucination
onecite benchmark --anti-hallucination --json

The repository baseline record is stored at benchmarks/leaderboard.json, with reproduction instructions in benchmarks/README.md.

Anti-hallucination evaluation

onecite benchmark --anti-hallucination runs a labelled, fully-offline evaluation of OneCite's core safety property. It resolves real strong identifiers (class A) into source-resolved BibTeX, leaves ambiguous plain-text references (class B) and fabricated, non-existent DOIs (class C — the kind a language model may hallucinate) unresolved rather than emitting a wrong citation, and flags mismatched pairings (class D — a real DOI attached to a different paper's title, the most common hallucinated-citation shape) with a text_metadata_mismatch warning instead of silently emitting them as clean source-resolved output. It reports three metrics:

• resolution rate — fraction of class-A inputs correctly resolved;
• non-fabrication rate — fraction of class-B/C inputs correctly left unresolved (not fabricated). 100% means OneCite invented no citations;
• mismatch detection rate — fraction of class-D inputs resolved with the mismatch warning attached.

A pipeline crash is recorded as error and never counts as correct for any metric — a clean rejection and a broken pipeline are different outcomes.

The dataset lives at src/onecite/benchmarks/anti_hallucination_cases.json, and the evaluation is also available from Python via onecite.run_anti_hallucination_eval().

onecite doctor

Check the local installation health for automation and CI. The doctor command checks package importability, bundled templates, packaged benchmark resources, the repository-contained OneCite Skill, and the offline benchmark regression check.

Usage:

onecite doctor
onecite doctor --json

The JSON output is a stable envelope with schema_version, tool, command, status, environment, summary, and checks fields.

OneCite Skill for Automated Workflows

The repository includes a local skill package at skills/onecite/SKILL.md. It gives automation and contributor workflows a repeatable procedure for reference cleanup, benchmark and doctor checks, and explicit reporting of unresolved entries. The skill is repository-contained and does not install itself into any local tool memory.

Input Type Auto-Detection

When --input-type is not specified, OneCite automatically detects the input type:

• Files ending with .bib are treated as BibTeX format
• All other files and strings are treated as plain text

Available Templates

OneCite supports several template presets for different entry types:

• journal_article_full - Full journal article entry (default)
• conference_paper - Conference proceedings paper
• book - Book entry
• thesis - Thesis/dissertation entry
• dataset - Dataset entry
• software - Software/code entry

Exit Codes

• 0 - Success
• 1 - Error occurred (invalid input, processing failure, etc.)
• 2 - One or more entries were unresolved when --fail-on-unresolved was used

For onecite benchmark and onecite doctor, exit code 0 means the configured checks passed and exit code 1 means at least one check failed.

🗺️ Roadmap

• OneCite Skill — Repository-contained operating guide for local citation-cleanup workflows
• Benchmarking — Small deterministic regression suite, configurable pass-rate gate, and baseline record
• Enhanced CLI — Automation-friendly JSON, NDJSON, summaries, and exit codes for reference processing
• Anti-hallucination evaluation — Labelled offline eval of the non-fabrication property (resolution, non-fabrication, and mismatch detection rates), gated in CI
• Audit-grade reports — Text/DOI mismatch warnings, failure reason codes with original input, DOI-level deduplication, and suggest source-health disclosure
• CSL-JSON output — --output-format csl-json emits CSL-JSON items for downstream tools that consume the format; a development fixture verifies Pandoc 3.10 consumption, while Quarto, standalone citeproc, and reference-manager imports are not separately validated
• Expanded suggest sources — Direct arXiv candidate search covers the CS venues that CrossRef does not index
• Concurrent batch resolution — Parallel source lookups for large reference lists (currently sequential; latency depends on the selected routes and external services)
• Larger anti-hallucination dataset — More labelled cases per class and a published live-mode baseline

🤝 Contributing

Contributions are always welcome! Please see CONTRIBUTING.md for development guidelines and instructions on how to submit a pull request.

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

OneCite