Repository-level retrieval-augmented generation (RAG) systems depend on vector indexes that must stay synchronized with code that changes continuously. The arXiv preprint 2610.25108 argues that there is currently no principled basis for deciding which changed code units actually need new embeddings, leaving index maintenance as an open problem.

To address this, the authors introduce CIPBench, a benchmark aimed at evaluating multi-signal index maintenance. The title points to "the limits of multi-signal index maintenance," suggesting the work examines how different signals can inform re-embedding decisions and where those signals fall short.

Note that the abstract is truncated, so this article cannot report specific findings, methodology details, or quantitative results. The source is a single preprint, and no comparison with other work is possible from the provided text.