SigMap TF-IDF hits 80% top-5

This is a strong narrow benchmark result, not proof that embeddings are obsolete. It does show that for offline local-model context compression, a cheap first-pass ranker can get much farther than many teams expect.

• –Function signatures and class shapes are unusually information-dense, so exact lexical matching has a real advantage over semantic paraphrase in codebases
• –The 98.1% token reduction is the practical headline: it makes local-model workflows cheaper and more repeatable before any vector stack is introduced
• –The likely ceiling is multi-hop and semantic queries, where naming alone stops being enough and call graphs or rerankers become necessary
• –The benchmark probably rewards well-named repos; generic helpers, deeply abstracted code, and cross-file flows will be the stress cases
• –For teams building lightweight code retrieval, this is a good case for “TF-IDF first, embeddings later” instead of starting with heavyweight RAG