To See Is Not to Master: LLMs Fail at Private-Library Code Generation Even With Accurate API Docs in Context
arXiv 2603.15159·medium signal
Researchers find that even when LLMs are given accurate, relevant private-library API documentation via RAG, they still fail to correctly invoke those APIs — showing that knowledge retrieval is necessary but not sufficient for private-library code generation. The gap is not in documentation access but in the model's ability to translate retrieved knowledge into correct invocation patterns. This has direct implications for any team using AI coding agents against internal APIs or proprietary SDKs.