PartsCleanse AI version 1.3 is now live in the AI2COE portal. The engine release introduces four significant improvements validated across five adversarial test catalogs and a 50,000-SKU benchmark.
TF-IDF weighted blocking replaces the previous keyword-block approach with a statistical relevance layer that improves both the accuracy of candidate pair selection and the computational efficiency of the comparison pipeline. At 50,000 SKUs, the engine completes in 42.9 seconds on standard hardware.
The part-type taxonomy has been expanded to 200+ categories covering bearings, valves, pumps, seals, motors, gaskets, instrumentation, electrical components, fasteners, hoses, piping, and structural parts families. The expanded taxonomy improves the discriminator penalty system's ability to distinguish between similarly described parts from different categories.
Manufacturer alias normalization now covers 800+ known name variants, abbreviations, and subsidiary relationships. The alias dictionary is augmented with a fuzzy fallback for manufacturer names not in the dictionary, reducing false mismatches from manufacturer naming inconsistency.
