citerra

BibTeX parser for Python.

citerra parses, validates, edits, and writes BibTeX documents. It supports strict parsing by default, opt-in tolerant recovery, diagnostics with source locations, raw-text retention, source-preserving writes, name/date/identifier helpers, and plain-record projection for application code.

The package is distributed as ABI3 wheels for Python 3.8 and newer.

Performance Snapshot

Measured on tests/fixtures/tugboat.bib: 2,701,551 bytes, 73,993 lines, and 3,644 entries. Hardware was AMD Ryzen 5 5600G, 6 cores / 12 threads. Measured on 2026-05-14 with Python 3.11.14; throughput is input-size normalized.

For a BibTeX parser, the relevant speed measurements are:

Workload	Why it matters	citerra result
Structured parse	Load a bibliography into entries and fields for application logic	0.008 s, 322.2 MiB/s
Source-preserving parse	Keep raw text, source locations, diagnostics, and source-order blocks for tools	0.011 s, 232.8 MiB/s
Raw-preserving write	Write retained source text after low-churn edits	0.002 s, 1149.8 MiB/s
Normalized write	Serialize structured data with configured formatting	0.010 s, 259.7 MiB/s

The comparison used citerra 0.3.0, bibtexparser 1.4.4, bibtexparser 2.0.0b9, and pybtex 0.26.1. citerra structured parse disables source capture and raw preservation for the closest parser-output comparison. Relative time is normalized to the first row in each table.

Python parser / mode	Version	Output retained	Median parse time	Throughput	Approx. entries/s	Relative time
citerra structured parse	0.3.0	Entries, fields, strings, comments, preambles	0.008 s	322.2 MiB/s	455.7k	1.0x
citerra source-preserving parse	0.3.0	Structured data, raw text, locations, diagnostics	0.011 s	232.8 MiB/s	329.3k	1.4x
bibtexparser parse	2.0.0b9	Entries/library model	0.367 s	7.0 MiB/s	9.9k	45.9x
pybtex parse	0.26.1	Bibliography data	0.863 s	3.0 MiB/s	4.2k	107.9x
bibtexparser parse	1.4.4	Entries/database model	10.758 s	0.2 MiB/s	0.34k	1345.2x

Python writer / mode	Version	Median write time	Throughput	Relative time
citerra raw-preserving write	0.3.0	0.002 s	1149.8 MiB/s	1.0x
citerra normalized write	0.3.0	0.010 s	259.7 MiB/s	4.4x
bibtexparser write	1.4.4	0.106 s	24.3 MiB/s	47.4x
bibtexparser write	2.0.0b9	0.497 s	5.2 MiB/s	222.0x
pybtex write	0.26.1	3.790 s	0.7 MiB/s	1691.3x

The workflow table below sums the median parse and write measurements for each parser/version. It is a round-trip estimate for parse-edit-write workloads, not a separate end-to-end benchmark.

Workflow	Median time	Throughput	Relative time
citerra source-preserving parse + raw-preserving write	0.013 s	193.6 MiB/s	1.0x
citerra structured parse + normalized write	0.018 s	143.8 MiB/s	1.3x
bibtexparser 2.0.0b9 parse + write	0.864 s	3.0 MiB/s	65.0x
pybtex 0.26.1 parse + write	4.653 s	0.6 MiB/s	349.7x
bibtexparser 1.4.4 parse + write	10.864 s	0.2 MiB/s	816.5x

Reproduction commands are listed in Reproducing Benchmarks.

Install

pip install citerra

The distribution name and import name are both citerra:

import citerra

Parse

import citerra

document = citerra.parse(
    '@article{paper, author = "Jane Doe", title = "Example Paper", year = 2026}',
    expand_values=True,
)

entry = document.entry("paper")
assert entry is not None
assert entry.entry_type == "article"
assert entry.get("title") == "Example Paper"
assert entry.date_parts().year == 2026

File helpers are available:

from pathlib import Path
import citerra

document = citerra.parse_path("references.bib", tolerant=True)
Path("normalized.bib").write_text(citerra.dumps(document), encoding="utf-8")

File-like helpers are also available:

with open("references.bib", encoding="utf-8") as handle:
    document = citerra.load(handle, tolerant=True)

text = citerra.dumps(document)

Document Model

• Document contains entries, comments, preambles, string definitions, source-order blocks, diagnostics, and validation helpers.
• Entry exposes the citation key, entry type, fields, source text, semantic helpers, and field mutation methods.
• Field exposes the original field name, parsed value, optional raw source text, and optional source location.
• Value represents string literals, numbers, variables, and concatenations.
• Diagnostic reports parse or validation problems with stable codes and source locations when available.

Tolerant Parsing And Diagnostics

text = '''
@article{ok, title = "Good"}
@article{bad, title = "Missing close"
@book{recovered, title = "Recovered"}
'''

document = citerra.parse(
    text,
    tolerant=True,
    capture_source=True,
    preserve_raw=True,
    source="refs/main.bib",
)

if document.status != "ok":
    for diagnostic in document.diagnostics:
        span = diagnostic.source
        if span is None:
            print(diagnostic.code, diagnostic.message)
        else:
            print(diagnostic.code, span.line, span.column, diagnostic.message)

Raw Text And Source-Preserving Writes

text = '@article{paper, title = "Example Paper"}'

document = citerra.parse(
    text,
    tolerant=True,
    capture_source=True,
    preserve_raw=True,
)

entry = document.entry("paper")
if entry is not None:
    print(entry.raw)
    print(entry.field("title").raw_value)

Use WriterConfig(preserve_raw=True) for low-churn output that reuses retained source text where possible. Use WriterConfig(preserve_raw=False) for normalized structured output.

document.rename_key("paper", "paper-v2")
document.set_field("paper-v2", "note", "accepted")
document.remove_export_fields(["abstract", "keywords"])

config = citerra.WriterConfig(
    preserve_raw=True,
    trailing_comma=True,
)
output = document.write(config)

Plain Records

Some application code wants ordinary dictionaries for filtering, indexing, or bulk transforms. citerra provides explicit helpers for that shape without changing the document model. Use write_entries() for new normalized output from records:

document = citerra.parse_path("references.bib")
records = citerra.document_to_dicts(document)

selected = [record for record in records if record.get("year") == "2026"]
text = citerra.write_entries(
    selected,
    field_order=["author", "title", "journal", "year", "doi"],
    sort_by=["ID"],
    trailing_comma=True,
)

Plain records use ENTRYTYPE and ID keys for the entry type and citation key. Values may be plain strings, integers, or Value objects.

For low-churn edits to a whole parsed document, overlay records back onto the document and then write the document. This preserves retained comments, preambles, string definitions, source order, and unchanged raw field text where possible:

document = citerra.parse_path("references.bib")
records = document.to_dicts()

for record in records:
    if record.get("doi"):
        record["note"] = "checked"

summary = document.update_from_dicts(records)
text = document.write()

Use value_mode="value" when applications need macro and concatenation-aware records:

records = document.to_dicts(value_mode="value")
records[0]["title"] = citerra.Value.from_bibtex_source('venue # " 2026"')
document.update_from_dicts(records)

Helpers

assert citerra.normalize_doi("https://doi.org/10.1000/XYZ.") == "10.1000/xyz"
assert citerra.latex_to_unicode("Jos\\'e") == "José"

names = citerra.parse_names("Jane Doe and {Research Group}")
assert names[1].literal == "Research Group"

date = citerra.parse_date("2026-05-13")
assert (date.year, date.month, date.day) == (2026, 5, 13)

Reproducing Benchmarks

The comparison script uses whichever optional packages are installed in the active environment. bibtexparser 1.x and 2.x use the same package name, so their rows are measured in separate environments.

python python/benchmarks/compare_parsers.py tests/fixtures/tugboat.bib --iterations 15 --warmups 3
python python/benchmarks/compare_parsers.py tests/fixtures/tugboat.bib --write --iterations 15 --warmups 3

Implementation

citerra is implemented as a native extension. Wheels include the parser engine, so ordinary Python installs do not require a Rust toolchain.

Rust Crate

The Rust crate is published as bibtex-parser on crates.io:

[dependencies]
bibtex-parser = "0.4"

See RUST.md for Rust usage.

Local Build

Use the project manifest for local development:

guix shell -m manifest.scm -- maturin build --release --out target/wheels

For local tests without installing into the user environment, unpack the built wheel into a temporary import directory and run pytest with that directory on PYTHONPATH:

rm -rf target/python-test
python3 - <<'PY'
from pathlib import Path
from zipfile import ZipFile

wheel = sorted(Path("target/wheels").glob("citerra-*.whl"))[-1]
target = Path("target/python-test")
target.mkdir(parents=True, exist_ok=True)
with ZipFile(wheel) as archive:
    archive.extractall(target)
PY
guix shell -m manifest.scm -- env PYTHONPATH=target/python-test python3 -m pytest tests/python

License

Licensed under either of Apache-2.0 or MIT, at your option.