Sparse (BM25) + dense (MedCPT) retrieval, fused and reranked with a MedCPT cross-encoder, over ~58k PubMed abstracts — with optional Gemini-generated, cited answers.
Precomputed from the real pipeline — loads instantly.
Unlike the examples above, nothing here is precomputed: your question goes through hybrid retrieval and MedCPT cross-encoder reranking of 100 candidates live, on a shared GPU (Hugging Face ZeroGPU). It typically takes about a second, or up to ~20 seconds when a GPU first has to be attached. If the demo hasn't been used in a while, the first search can take a few minutes while the ZeroGPU Space starts up.