Show HN: Semble – Code search for agents that uses 98% fewer tokens than grep

MinishLab open-sourced Semble, a code-search tool for agents that combines Model2Vec embeddings, BM25, RRF fusion, and reranking; on a 63-repo benchmark, it used 98% fewer tokens than grep+read, reached 0.854 NDCG@10, and ran CPU queries in about 1.5 ms.

Why it matters: HKR-H/K/R all pass: the 98% token claim is clickworthy, the 63-repo benchmark adds substance, and coding-agent context cost is a real practitioner nerve. Impact is still toolchain-level, so it stays below must-write.