This tool uses Xenova's uncased DistilBERT model fine-tuned on SQuAD 1.1 through Transformers.js. It divides longer context into overlapping passages, reads each passage in a browser worker, and ranks the returned spans. WebGPU is tried first, with a WebAssembly CPU fallback when graphics acceleration is unavailable.
Extractive question answering does not write a new explanation or search the web. It selects words already present in the supplied context. The first run downloads public runtime, tokenizer, and model files that total tens of megabytes. Browsers normally cache them, but cleared storage or private browsing can cause another download.