Type or say your question. The assistant searches its knowledge base with keyword and embedding search, re-ranks the results with a trained model, and quotes the answer word for word with its source. If the answer isn't there, it tells you instead of guessing.
SQuAD v1.1 dev questions, none used for training. Each has a known source passage, so every number below is checked against the truth.
by retriever
| Retriever | Recall@1 | @5 | @10 | MRR@10 | p95 |
|---|
Recall@k is the share of questions whose source passage is in the top k results. Latency is per query on a 2-vCPU machine.
Recall@1
Everything runs on CPU inside one serverless function. The highlighted steps are models trained for this project.
Two searches, one ranking. Keyword search (BM25) and embedding search each pull their top 50 passages. A small neural ranker, trained on 15,000 separate questions, orders the merged list using 16 signals: both scores, how many of the question's rare words a passage contains, how tightly they cluster in one or two sentences, and title overlap.
Quotes, not paraphrases. A second model picks the sentence most likely to hold the answer from the top three passages, so every answer is a verbatim quote with a citation. A confidence model, trained on questions about articles left out of the index, decides when to decline. Voice input uses your browser's own speech recognition, and read-aloud uses its built-in voices.