Extractive QA: a different tool for a different job
Paste a paragraph of context, ask a factual question, and the model finds the precise answer inside that text and highlights it. The DistilBERT model (fine-tuned on SQuAD 2.0) reads your context and question together, then points at the span that answers it. The defining feature — and the reason to trust it — is that it can only return words you provided. There’s no outside knowledge to be wrong about and no creative generation to hallucinate. The answer is always traceable to a spot in your source.
What it’s great at — and what it isn’t
| Works well | Doesn’t work |
|---|---|
| ”When was the contract signed?" | "Is this a good contract?” (opinion) |
| “How many users did they report?" | "Summarize this” (use the summarizer) |
| “Who is the CEO?” | Questions needing outside facts |
| Factual who/what/when/where/how-many | Inference across many paragraphs |
It answers questions with a direct, stated answer in the text. Questions requiring judgment, synthesis across the whole document, or knowledge beyond the context are the wrong fit.
Handling long documents
The model processes about 512 tokens (~300–400 words) of context at once. For longer texts, the tool splits the context into overlapping chunks and finds the best answer across them — but an answer that’s stated in one place and referenced in another can still slip through the chunk boundaries. If you know roughly where the answer lives, pasting just that section gives the cleanest result.
Where it shines
This is the right instrument for grounded answers — pulling a specific figure from a report, finding a clause in a contract, answering a comprehension question about a passage — where you need to see exactly where the answer came from. It runs entirely in your browser, so the documents you query never leave your device. For open-ended “explain or reason about this” tasks, a generative assistant is the better fit; for “find the answer in this text, and show me where,” extractive QA is purpose-built.