How Hugging Face Inference Endpoints, Jobs, and Buckets Power Search on Papers with Code

The revived Papers with Code search separates preparing its research catalog from answering a reader's query. In the Hugging Face engineering account, batch Jobs convert paper titles and abstracts into numerical representations of meaning, with Buckets holding the resulting artifacts. An online Inference Endpoint creates the corresponding representation for each search query. Keeping both sides on the same versioned format makes their comparison consistent.

Search combines meaning-based matches with ordinary text matches. If the model endpoint is unavailable or exceeds a one-second timeout, the application returns text-search results. That fallback keeps search usable when the model service is starting up or busy, while changing which retrieval method supplies the results.