Native reranking-model support in Ollama, so a local RAG pipeline doesn't need a second inference server
Self-hosters running local RAG standardize on Ollama for LLM and embedding serving, then hit a wall: Ollama has no reranking endpoint, so pipelines must spin up a separate inference server (HuggingFace TEI, Infinity, or llama.cpp) just for the cross-encoder rerank step. The two-year-old request is one of Ollama's most-reacted issues and is still hot -- users are publishing hacky workarounds in June 2026 because it's 'not currently supported.' A native rerank endpoint would let an all-Ollama local-RAG stack run on one binary.
The build isn't a model, it's the missing /rerank endpoint and library entries so Open WebUI / LangChain pipelines 'just work' against one Ollama; whoever lands the cross-encoder serving path (or a clean Ollama-compatible sidecar) inherits two years of pent-up local-RAG demand.
landscape (2 existing solutions)
Local-RAG builders standardize on Ollama for LLM + embeddings, then have to bolt on a separate reranker server; a native Ollama rerank endpoint plus model-library entries would close the last gap in an all-Ollama pipeline.