srdedupe 0.1.1

Safe, FPR-controlled, reproducible deduplication pipeline for bibliographic reference files, designed for systematic review workflows.

srdedupe — Safe Bibliographic Deduplication

Safe, reproducible deduplication for systematic reviews and bibliographic databases.

Parses and deduplicates bibliographic reference files (RIS, NBIB, BibTeX, WoS, EndNote) with FPR-controlled decision making, full audit trails, and deterministic outputs.

Installation

pip install srdedupe

Quick Start

Parse and export

from srdedupe import parse_file, parse_folder, write_jsonl

# Single file (format auto-detected)
records = parse_file("references.ris")

# Entire folder
records = parse_folder("data/", recursive=True)

# Export to JSONL
write_jsonl(records, "output.jsonl")

Deduplicate

from srdedupe import dedupe

result = dedupe("references.ris", output_dir="out", fpr_alpha=0.01)

print(f"Records: {result.total_records}")
print(f"Auto-merged clusters: {result.total_duplicates_auto}")
print(f"Review records: {result.total_review_records}")
print(f"Unique records: {result.total_unique_records}")
print(f"Dedup rate: {result.dedup_rate:.1%}")
print(f"Output: {result.output_files['deduplicated_ris']}")

CLI

# Parse to JSONL
srdedupe parse references.ris -o output.jsonl
srdedupe parse data/ -o records.jsonl --recursive

# Full deduplication pipeline
srdedupe deduplicate references.ris
srdedupe deduplicate data/ -o results --fpr-alpha 0.005 --verbose

How It Works

A 6-stage pipeline controlled by false positive rate (FPR):

1. Parse & Normalize — Multi-format ingestion, field normalization
2. Candidate Generation — High-recall blocking (DOI, PMID, year+title, LSH)
3. Probabilistic Scoring — Fellegi-Sunter model with field-level comparisons
4. Three-Way Decision — AUTO_DUP / REVIEW / AUTO_KEEP with Neyman-Pearson FPR control
5. Global Clustering — Connected components with anti-transitivity checks
6. Canonical Merge — Deterministic survivor selection and field merging

Pairs classified as REVIEW are preserved in output artifacts for manual inspection.

API Reference

parse_file(path, *, strict=True) -> list[CanonicalRecord]

• Parse a single bibliographic file. Format is auto-detected from file content.

parse_folder(path, *, pattern=None, recursive=False, strict=False) -> list[CanonicalRecord]

• Parse all supported files in a folder. Optional glob pattern (e.g. "*.ris").

write_jsonl(records, path, *, sort_keys=True) -> None

• Write records to JSONL file with deterministic field ordering.

dedupe(input_path, *, output_dir="out", fpr_alpha=0.01, t_low=0.3, t_high=None) -> PipelineResult

Run the full deduplication pipeline. Returns a PipelineResult with:

• success, total_records, total_candidates, total_duplicates_auto, total_review_records, total_unique_records, dedup_rate
• output_files — dict mapping artifact names to file paths
• error_message — error details if success is False

Advanced: PipelineConfig + run_pipeline

For full control (custom blockers, FS model path, audit logger):

from pathlib import Path
from srdedupe.engine import PipelineConfig, run_pipeline

config = PipelineConfig(
    fpr_alpha=0.01,
    t_low=0.3,
    t_high=None,
    candidate_blockers=["doi", "pmid", "year_title"],
    output_dir=Path("out"),
)

result = run_pipeline(input_path=Path("references.ris"), config=config)

Supported Formats

Format	Extensions
RIS	.ris
PubMed/NBIB	.nbib, .txt
BibTeX	.bib
Web of Science	.ciw
EndNote Tagged	.enw

Pipeline Output Structure

out/
├── stage1/canonical_records.jsonl
├── stage2/candidate_pairs.jsonl
├── stage3/scored_pairs.jsonl
├── stage4/pair_decisions.jsonl
├── stage5/clusters.jsonl
├── artifacts/
│   ├── deduped_auto.ris
│   ├── merged_records.jsonl
│   ├── clusters_enriched.jsonl
│   ├── review_pending.ris  (if review pairs exist)
│   └── singletons.ris      (if singletons exist)
└── reports/
    ├── ingestion_report.json  (folder input only)
    └── merge_summary.json

Development

make dev           # Install dependencies + pre-commit hooks
make test-fast     # Quick validation while coding
make check         # Lint + type check + format (before committing)
make test          # Full test suite (417 tests, ≥80% coverage)

Documentation

• CONTRIBUTING.md — Code style, testing, contribution guidelines

License

MIT — see LICENSE.

Citation

@software{srdedupe2026,
  author = {Lopes, Ennio Politi},
  title = {srdedupe: Safe Bibliographic Deduplication},
  year = {2026},
  url = {https://github.com/enniolopes/srdedupe}
}