Code Search

Find the code that does this thing, in a codebase nobody has read all of. It looks like a text retrieval problem and is not quite one: code has structure that prose does not — an identifier is defined exactly once, a call graph is a real graph, and a type says what a function accepts. Retrieval that ignores that is throwing away the part of the corpus that is unambiguous.

The two failure modes pull opposite ways. Lexical search finds the exact identifier and misses the function that does the same job under another name; embedding search finds the paraphrase and returns something plausible from the wrong module. What works is neither alone — structure narrows the candidate set to what could possibly be relevant, and semantics ranks within it, which is a different pipeline from the hybrid retrieval used on prose.

The share of queries that structure alone can answer is the useful quantity, because those are the ones that can be answered exactly rather than probably. A query naming an identifier, a type or a call site is resolvable by a symbol index with no model involved, and it is a large share of real queries. Semantic retrieval is then working on the remainder — which is the harder remainder, and a much smaller one than an evaluation on paraphrase queries suggests.