citation-integrity-check 0.1.0

Verify answer citations refer to supplied source ids and that cited sources actually support the claims. Python port of @mukundakatta/citation-integrity-check.

citation-integrity-check

Verify answer citations refer to supplied source ids and that cited sources actually support the claims. Zero runtime dependencies.

Python port of @mukundakatta/citation-integrity-check. The JS sibling has the original API; this README sticks to the Python surface.

Install

pip install citation-integrity-check

Usage

from citation_integrity_check import verify

sources = [
    {"id": "1", "text": "Photosynthesis converts light into chemical energy in plants."},
    {"id": "abc123", "text": "Chlorophyll absorbs red and blue wavelengths of light."},
]
answer = (
    "Plants use photosynthesis to convert light into energy [1]. "
    "Chlorophyll absorbs red and blue light [id:abc123]."
)

result = verify(answer, sources)

result.ok            # True if no missing ids and no unsupported claims
result.missing       # list[str]    -- cited ids that don't exist in sources
result.unsupported   # list[Claim]  -- sentences with no valid supporting citation
result.coverage      # float in [0, 1] -- fraction of sentences with a valid citation

Citation forms

Two markers are recognized inside the answer:

Form	Resolves to
[1]	sources[0] (1-based index) and source.id == "1"
[id:abc]	source.id == "abc"

Anything else inside brackets (like [Note]) is ignored, so stylistic prose doesn't count as a citation.

How "unsupported" is decided

A sentence is unsupported when any of these is true:

• It has no citation marker at all (reason="no_citation").
• All cited ids are missing from sources (reason="missing_source").
• The cited source's text doesn't share enough non-stopword tokens with the sentence (reason="insufficient_overlap").

Token-overlap is |claim_tokens & source_tokens| / |claim_tokens|, with a small built-in stopword list. The threshold is tunable:

verify(answer, sources, support_threshold=0.5)  # stricter

API differences from the JS sibling

• Returns a CitationResult dataclass with unsupported claims (per-sentence) instead of the JS unused ids list.
• Adds the [id:foo] named-citation form alongside numeric [N].
• Adds the token-overlap support_threshold to verify the cited source actually mentions the claim.

See the JS sibling's README for the full design notes.