Ask your documents

• Small offices with handbooks, contracts and procedures that must not leave the building.

Handbooks, procedures and supplier terms hold the answer to everyday questions, but finding the right paragraph takes time. An answer without its source is hard to check, and many offices cannot send internal documents to a cloud service.

1. Reads a folder of PDF (with a text layer), Markdown or text files, and cuts them into short passages that keep their file name, page and nearest heading.
2. Searches with BM25, a standard keyword-ranking formula. No AI model, no embeddings, no network.
3. Tries to refuse unsupported questions, under defined rules. If a name in the question appears in no document, or the best passage does not contain at least 60% of the question's key terms, the answer is "Not found in the documents.", with the missing terms listed. The rules are not perfect: see the false answer in the results below.
4. Answers with the best-matching sentence, copied word for word, with its file, page and section. Every answer and every refusal shows a plain reason.

Optional: an Ollama model can reword the answer. It is off by default and takes no part in the search or the "not found" decision. When it is on, the question and the retrieved passage are sent to the Ollama address in the OLLAMA_HOST setting: by default this computer (localhost), but if that setting points to another machine, they leave this one. The rewording is model output and is labelled as such; wording that adds numbers or names not in the passage or the question is not shown, and the cited quote is always shown next to it.

8 invented documents about an invented company (68 passages), with questions whose answers are known, including questions the documents do not answer. Per the developer's record, settings were tuned on the development set only and the hold-out questions were never used for tuning, so the hold-out is the fairer number. Both sets ask about the same 8 documents; the hold-out holds back questions, not documents. Measured 5 October 2026; figures from the project's evaluation summary.

• The baseline uses the same search with the name and 60% refusal rules switched off; it still says "not found" when no passage shares any searchable word with the question. On the hold-out set it gave a wrong answer to all 12 unanswerable questions; with the "not found" rules, 11 of 12 were refused correctly. With only 12 such questions, all about the same 8 documents, this count is not a reliable rate.
• The trade-off: 5 answerable hold-out questions were wrongly marked "not found", so correct answers dropped from 27 of 31 to 25 of 31.
• One false answer remains: asked "Who is the safety officer?", the tool returned the rule that mentions the safety officer.
• These are small synthetic sets (85 questions over 8 documents) and do not establish accuracy on your documents.

Python, pypdf, snowballstemmer (word stems for search), pandas, Streamlit; tests with pytest and Playwright; packaged with Docker.

• Word matching, not understanding. A question that uses other words than the document ("USB drives" against "USB storage devices") can get "not found".
• Near-miss questions about a detail next to a covered topic can still get a wrong answer.
• Scanned PDFs need OCR (text recognition), which is not included. Only PDFs with a text layer are read. Tables lose their layout.
• One language at a time. An English question does not find an Italian passage, and the other way round.
• Input limits: files over 20 MB, uploads over 100 MB or 200 files, and PDFs over 500 pages are refused. A hostile PDF within these limits can still be slow.
• Single-user demo: document indexes are kept in the app's memory for up to an hour and are shared by everyone using the same server, so it should not be exposed to other users without authentication.
• Not legal advice: always read the cited passage in the original document.

Source: snello-docqa README and reports/summary.json and eval_holdout.json, measured 5 October 2026 at commit fc6bcbe (measured on 5 October 2026 at commit fc6bcbe). Code is private, so it is not linked.