Keeping it boring (and relevant) with BM25F

The adaptation of the BM25 algorithm to improve code search efficiency by 20% at Sourcegraph highlights the enduring relevance of classic search methodologies in modern applications. BM25, originally designed for text search, has been tailored to handle code-specific complexities, such as distinguishing between matches in file names and symbol definitions versus less significant matches in comments or statements. This adaptation involved the integration of BM25F, an extension that allows for differential weighting of term matches across different fields like content, symbols, and filenames, ensuring a balanced and interpretable ranking system. The revised approach enhances Sourcegraph's search capabilities by leveraging line-level scoring and combining traditional keyword search with semantic methods to address diverse customer needs, including support for niche programming languages and proprietary systems. Despite advancements in search technology, the foundational strength of BM25, with its simplicity and conceptual rigor, continues to prove valuable, as demonstrated through rigorous internal evaluations against modern embeddings and semantic reranking models.