tai/research/fuzzy-matching-selectors/github-rapidfuzz-fuzz-py.md

RapidFuzz fuzz_py.py (WRatio, partial_ratio)

Metadata

Field	Value
Institute	Molodetz Institute — Research & Retrieval Division
Subject	fuzzy-matching-selectors
Document type	Source Dossier
Tier	1
Source title	RapidFuzz fuzz_py.py (WRatio, partial_ratio)
Source URL	https://raw.githubusercontent.com/rapidfuzz/RapidFuzz/main/src/rapidfuzz/fuzz_py.py
Source publisher / author	maxbachmann and RapidFuzz contributors
Source type	repository
Source date (published)	Unknown — not stated
Date researched	2026-10-25
Researched by	tai (agent), model: Claude Sonnet 5, on behalf of retoor
Research method	direct page fetch
Confidence	High
Status	Active

Summary

The pure-Python reference implementation of RapidFuzz scorers, read for how the weighted WRatio combines ratio, token and partial scorers.

Key findings

• WRatio constants: UNBASE_SCALE 0.95; PARTIAL_SCALE 0.9 when the length ratio is <= 8.0, else 0.6.
• Length ratio < 1.5: max(ratio, token_ratio * 0.95) where token_ratio is the max of token_set_ratio and token_sort_ratio.
• Length ratio >= 1.5: ratio, then partial_ratio * PARTIAL_SCALE, then partial_token_ratio * 0.95 * PARTIAL_SCALE; score_cutoff is raised after each step so later scorers can prune.
• partial_ratio slides the shorter string over the longer string (prefix windows, full-length windows, suffix windows), scoring each with normalized Indel similarity and keeping the best.
• Empty input or a None input returns 0.

Pros / strengths

Precise, small and readable; shows the max-of-scaled-scorers pattern tai also uses.

Cons / limitations / caveats

Designed for sentences and records, not identifiers. The docstring and code disagree for needles longer than 64 characters (noted by the fetch).

Relevance to tai

Validated the 'max of several scaled signals' structure. Not ported: token_set semantics (a subset scores 100) would make claude tie with every claude-* id and hide real differences.

Changelog

Date	Change
2026-10-25	Initial dossier created.