The llama.cpp b11818 release adds support for typed content input to its server’s rerank endpoint. The update also rejects multimodal rerank requests when the model lacks a rerank chat template, rather than silently dropping images. Invalid media downloads or base64 data, and negative `top_n` values, now receive a 400 error instead of a server error.
The README also says the TEI input accepts the same shape as Jina. The release lists builds for several operating systems and hardware configurations; the source does not specify a release date beyond its publication timestamp.
Why it matters
Developers using llama.cpp’s rerank endpoint can send typed content, while callers using models without a rerank chat template get an explicit rejection instead of having images silently ignored.
