@aigclink: An Agent-oriented code search tool: Semble. It uses natural language to search codebases and returns precise code snippets, saving 98% token consumption compared to grep+read. The method lets Agents use natural language to directly locate the most relevant lines of code, without guessing keywords or reading entire files. Speed: indexing a typical…

An Agent-oriented code search tool: Semble. It uses natural language to search codebases and returns precise code snippets, saving 98% token consumption compared to grep+read. The approach lets Agents use natural language to directly locate the most relevant lines of code, without guessing keywords or reading entire files. In terms of speed, indexing a typical repository takes about 250 milliseconds, and answering a query takes about 1.5 milliseconds, with NDCG@10=0.854. Semble achieves 94% recall with 2k tokens, while grep+read only reaches 85% with a 100k token window. First, it uses intelligent chunking: not mechanically cutting by line count, but intelligently splitting by code structure (functions, classes, code blocks), making each chunk a meaningful code unit without cutting a function in half. The search returns semantically complete code blocks — a function, a class, or an independent logic unit — without truncating code. Semantic and lexical dual-path retrieval + Reciprocal Rank Fusion (RRF) fusion mechanism: each path produces a ranking, and RRF merges them. Code-aware re-ranking: after initial results, the rankings are fine-tuned using code-specific signals to push the most relevant ones to the top. It should be quite effective for Agent code search scenarios. #Semble #AgentCodeSearch