Less noise.
Better evidence.
Rerank search results and see the passage behind each score.
Your documents stay in this browser.
Inspect the evidence
Ranked by relevance, with the strongest passage from each document.
A first-place result is not proof of an answer. This model ranks candidates even when none is relevant.
The useful part, brought forward.
Choose an example or paste your own candidates.
Then run the model to see what belongs first.
A second look at your search results.
Pocket Rerank scores question–passage pairs. Overlapping windows check beyond the beginning of each document. The highest-scoring window becomes its evidence.
Evidence text is decoded from tokens and may normalize spacing. Browser windows use direct token slices; Python uses original character offsets. Their long-document behavior is not identical.
Useful measurements. Clear limits.
The Python runtime scored 0.721 NDCG@10 on 300 SciFact queries, versus 0.652 for keyword search. The advantage over truncated reranking (0.716) was inconclusive.
This is one dataset, not a general accuracy claim. Browser speed depends on your device. Scores indicate relevance, not truth or probability. English only; long documents may receive higher scores.
Read the evaluation reports ↗Limits, privacy & attribution
Browser limits: 16 documents, 12,000 characters per document, 48,000 total characters, 128 query tokens and 64 windows. Oversized inputs fail explicitly. No input text is sent to an inference service. Hugging Face receives model download requests. No analytics are used by this app. The model runs in a worker so you can stop it.
Converted from Ettin 17M (Apache 2.0). Original weights belong to their upstream authors. Browser runtime: Transformers.js and ONNX Runtime Web. This is not a newly trained foundation model.