AI Question Answering — Free Extractive QA Over Any Passage

The tool runs two analyses on every question and shows them side-by-side, so you can sanity-check the model against a transparent baseline. Inference happens on the server — no model weights are downloaded into your browser, and the page works on any device that can hit our API.

Neural answer. The passage and question are sent to the Hugging Face Inference API for distilbert/distilbert-base-cased-distilled-squad, a DistilBERT (cased) checkpoint fine-tuned on SQuAD 1.1. The model attends to the passage and the question together, then emits two probability vectors of length 512 — start logits and end logits. The library scores every valid (start, end) pair where start ≤ end and the span length is within the configured cap, then returns the highest-scoring span and its softmax-normalised probability as the "confidence".

Long passages. The DistilBERT context window is 512 sub-word tokens. When your passage exceeds that, the pipeline applies a sliding window of 384 context tokens with a stride of 128 tokens, runs inference on every chunk, then deduplicates overlapping spans and keeps the best one. The result panel shows the chunk count so you can see when this kicks in.

Lexical baseline.Independently of the model, a tiny keyword-overlap scorer runs over your passage. For each sentence it counts the fraction of content words (non stop-words) from your question it contains. A small bonus is added when the sentence matches the question's wh-pattern — numbers and durations for "how long / how many", calendar terms for "when", and capitalised noun phrases for "who / where". The top sentence is returned with a 0–1 score.

Cross-check. The two answer spans are compared with the official SQuAD word-overlap F1 metric (the same one the SQuAD leaderboard reports). High F1 means the two methods picked essentially the same words; low F1 means the model disagreed with the baseline and the answer is worth a second look. Both spans, both confidences, and the F1 are shown in the result tiles.

Confidence bands.The reported model confidence is calibrated to the model card's dev-set behaviour: ≥ 80% is high (usually correct on clean factoid questions), 50–80% is medium (treat as a candidate), and below 50% is low (the passage probably does not state the answer outright). Both the F1 cross-check and a low-confidence warning render automatically when the score drops into the low band.

Privacy. Text is sent once for scoring and not stored. No tracking, no ads, no signup. If you need to keep the passage local, the same model can be run via transformers in Python on your own machine; see the sources below.

• Rajpurkar et al. (2016) — SQuAD: 100,000+ Questions for Machine Comprehension of Text
• Sanh et al. (2019) — DistilBERT: smaller, faster, cheaper and lighter
• Hugging Face model card — distilbert-base-cased-distilled-squad
• Hugging Face Inference API — server-side inference endpoint
• SQuAD evaluation conventions — exact-match and F1 scoring

Cross-checked against the official model card and SQuAD reference scorer on 2026-05-12. The page is reviewed when the upstream checkpoint changes or when a clearly better extractive QA backbone (XLM-RoBERTa-SQuAD2 for multilingual, DeBERTa-v3-SQuAD2 for English with no-answer support) becomes available on the free Inference API tier.