How to use
- Press Download model once (both models, about 89 MB).
- Add PDF, Word or text files, or paste text; they are split into passages and indexed on your device when you ask.
- Ask a question; the answer is highlighted in the passage it came from, with the document, page and confidence.
Worked example
Asking “How long is the notice period?” of the sample tenancy agreement returns “two months”, highlighted in the clause about written notice. Phrasing matters: if an answer looks off, ask more specifically.
Supported formats and limits
| Input | PDF (with a text layer), DOCX, TXT, MD, Pasted text |
|---|---|
| Output | Answers quoted from your documents, with the source passage, page and confidence |
| Limits | The first use downloads about 89 MB of model files (MiniLM search 23 MB + DistilBERT answer extraction 66 MB) from Hugging Face, cached afterwards. Documents never leave this tab. |
| Engine | all-MiniLM-L6-v2 sentence embeddings (Apache-2.0) to rank passages of about 60 words, then DistilBERT base cased distilled SQuAD (Apache-2.0) extracts the answer span from the best 4 passages; both via transformers.js in a Web Worker |
Limitations
- Answers are short spans copied from one passage; the model cannot combine facts from different places or answer yes/no questions well.
- If no span is confident enough, the tool says it found no answer rather than guessing.
- Scanned PDFs need OCR first. English only.
Questions
How does it find answers?
all-MiniLM-L6-v2 ranks passages of about 60 words by similarity to your question, then DistilBERT (SQuAD) copies the most likely answer span from the top 4 passages. Both are Apache-2.0 models, about 89 MB together, downloaded once from Hugging Face.
Is it like a chatbot?
No. It does not write new text; it extracts a short span from one passage. It cannot combine facts from different places and handles yes/no questions poorly. If nothing is confident enough, it says it found no answer instead of guessing.
Are my documents uploaded?
No. Documents are split and indexed in this tab. Scanned PDFs need OCR first, and the models are trained for English.
Privacy
On-device model. Processing runs in this browser. The open-source model files are downloaded once from the model host (Hugging Face) and cached; your content is not uploaded.
- Hugging Face: Only after you press Download model: requests for the model files (all-MiniLM-L6-v2, about 23 MB, and DistilBERT SQuAD, about 66 MB) from huggingface.co and its CDN, which see your IP address. Your text and files are never sent.
See the privacy policy for how toolsdocks handles data.