MediaWiki Code2Code Search: Neural Retrieval for the Semantic Discovery of Open-Source Software Entities

Researchers introduced MediaWiki Code2Code Search, a neural code-to-code retrieval system for semantic search across MediaWiki repositories.

It indexes 1.29 million structural entities from more than 2,500 repositories using a split GPU/CPU pipeline and a compact FAISS IVF-PQ index.

On a 27-query benchmark, it outperformed a BM25 baseline, with the biggest gains on name-obfuscated queries.

The result shows semantic retrieval can improve large-scale code/entity search while remaining fast and memory-efficient enough for real deployment.