Spaces:
Running
Apply for a GPU community grant: Personal project
ResearchIT is a free arXiv discovery app I'm building as a side project: a personalised feed that keeps a reader's separate research interests apart instead of blending them into one, plus hybrid search over ~1.97M arXiv papers (BGE-M3 dense vectors + SQLite FTS5 keyword search, fused with RRF, then a cross-encoder reranker). No account or sign-up needed.
Open parts: the Space's source is public, the 1.97M-paper metadata + citation sidecar it searches is a public dataset (siddhm11/researchit-metadata), and it runs on open models (BAAI/bge-m3, mixedbread-ai/mxbai-rerank-xsmall-v1).
Why a GPU: on 2 vCPU the cross-encoder is the slowest step, about 0.75 s of a ~1.4 s search, and that is with only 10 candidates and abstracts cut to ~500 characters. On our evaluation set (847 known-item queries, 487 from LitSearch), reranking with full abstracts raises MRR@10 from 0.563 to 0.611, and to 0.618 when reranking 20 candidates. On CPU a single full abstract roughly doubles reranking time (0.50 s to 1.22 s measured on this Space), so it stays off. On a GPU we can turn it on and keep search well under a second.
Request: Nvidia T4 small. ZeroGPU isn't an option because this is a Docker Space (FastAPI + HTMX), not Gradio. The code picks its device at startup, so the Space keeps working on CPU if the grant ends.