Atlas research paper search

Atlas is a research-paper search engine for people who work on recommender systems, ranking, and information retrieval. It began as the knowledge-search capability behind aipoweredrecsys.com: a way to find papers that matter to practitioners, without treating personal notes or a private knowledge base as the corpus.

The public experience is a free preview. Signed-in visitors get a limited number of searches. Results return bibliographic metadata—title, authors, venue, year when available—and a link to the canonical source. Atlas does not rehost paper PDFs or expose page images, excerpts, or generated summaries through the public API.

Under the hood, Atlas owns ingestion, rights assessment, indexing, retrieval, and a versioned search API. The website owns identity, rate limits, and UI. The current serving cascade encodes each query with ColQwen2, shortlists papers by fusing BM25 over page text with compact pooled visual vectors, then reranks candidate papers with late-interaction MaxSim before projecting rights-safe metadata. The corpus started from book-chapter references and an approved arXiv shortlist; the indexed thin slice used for the preview contains tens of papers and hundreds of pages rather than a comprehensive literature graph.

I am the sole human contributor and built Atlas with assistance from Cursor: rights-aware contracts that separate private indexing from public display, the hybrid multimodal cascade, a small repeatable smoke evaluation, and a Modal-hosted FastAPI endpoint consumed by the site. The preview is live and deliberately cost-constrained—the GPU service scales to zero when idle, so the first search after a quiet period can take up to about a minute.

What is implemented today is a working papers thin slice plus an integrated public preview. Broader evaluation, larger corpus expansion, additional source types such as GitHub and video, and quality tracks beyond the smoke harness remain planned or in progress rather than finished product claims.