Reference Finder

Source: https://github.com/aipoch/medical-research-skills

When to Use

• You have a scientific paragraph and want suggested PubMed papers for each sentence.
• You need top-ranked references with title, DOI, PMID, year, and a short why recommended explanation.
• You are drafting or reviewing a manuscript and want quick literature grounding for key claims.
• You want a lightweight reference matcher that uses only the official PubMed E-utilities API (no third-party services).
• You need a scriptable tool for batch or CLI workflows to generate candidate citations.

Key Features

• Sentence-level reference matching for scientific text.
• Returns the top N (default: 3) most relevant PubMed records per sentence.
• Outputs structured fields: title, DOI, PMID, year, recommendation reason.
• Relevance ranking based on:
  • keyword overlap / match strength,
  • publication year preference,
  • citation-count signal (when available/derivable).
• Safety constraints:
  • Network access restricted to eutils.ncbi.nlm.nih.gov.
  • No local filesystem writes except to outputs/ during execution.
  • Request timeout set to 30 seconds with clear error messages.
• Supports Python API usage and CLI usage (including interactive mode).

Dependencies

• Python 3.x (standard library only; no third-party packages required)

Example Usage

Python (direct call)

from reference_finder import find_references

text = "CRISPR-Cas9 gene editing has revolutionized biomedical research."

results = find_references(text)

for ref in results[:3]:
    print(f"- {ref['title']} ({ref['year']})")
    print(f"  DOI: {ref['doi']}")
    print(f"  PMID: {ref['pmid']}")
    print(f"  Reason: {ref['reason']}")

CLI (single input)

python scripts/find_refs.py "CRISPR-Cas9 gene editing has revolutionized biomedical research."

CLI (interactive mode)

python scripts/find_refs.py

Example output (JSON)

[
  {
    "pmid": "PMID:",
    "title": "A Programmable Dual-RNA-Guided DNA Endonuclease in Vitro",
    "doi": "10.1126/science.1225829",
    "year": 2012,
    "reason": "Highest keyword match for 'CRISPR-Cas9', foundational paper"
  }
]

Implementation Details

Data flow

1. Sentence splitting: The input text is split into sentences (implementation-defined; typically punctuation-based).
2. PubMed search (ESearch): For each sentence, a query is sent to:
   • https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi
3. Record retrieval (EFetch): The top candidate PMIDs are fetched via:
   • https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi
4. Field extraction: Title, year, PMID, and DOI (when present) are extracted from the returned metadata.
5. Ranking and selection: Candidates are scored and the top N are returned with a short recommendation reason.

Ranking signals

• Keyword match: Measures overlap between sentence terms and retrieved record metadata (e.g., title/abstract terms when available).
• Publication year: Used as a preference signal (e.g., favoring more recent work unless a classic/foundational match is strong).
• Citation count: Incorporated when available/derivable; otherwise treated as missing without failing the run.

Operational constraints and safety

• Allowed network host: eutils.ncbi.nlm.nih.gov only.
• Prohibited: Any third-party URLs.
• Filesystem: Do not write outside outputs/ during execution.
• Rate limiting: Use a reasonable request cadence (e.g., ~0.5s between requests) to respect API limits.
• Timeout: 30 seconds per request.
• Error handling: Return semantic, user-readable error messages for network/API/parse failures.

Defaults

• Top references per sentence: 3
• Endpoints:
  • ESearch: https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi
  • EFetch: https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi

Related project files

• Main script: scripts/find_refs.py
• Tests: tests/test_finder.py
• Evaluation checklist: references/evaluation-checklist.md
• PubMed E-utilities documentation: https://www.ncbi.nlm.nih.gov/books/NBK25504/