How it works
Traceable answers, not confident guesses
DocuLens answers questions about a document you provide and shows the evidence behind every answer, so you can check it instead of trusting it.
- 01
Validate
Extension and actual file content are checked (a .pdf must start with a PDF header; text must be valid UTF-8). Files over 4 MB or 200 pages are rejected.
- 02
Extract
PDF text is extracted on the server page by page; text files are decoded safely. Image-only PDFs are reported — there is no OCR.
- 03
Chunk
Text is split into passages of about 650 characters along sentence and paragraph boundaries. Each passage keeps its page number (PDF) or line numbers (text) and exact character offsets.
- 04
Retrieve
Questions are matched against passages with BM25, a transparent keyword-ranking method. If no passage contains enough of the question's key terms, DocuLens reports insufficient evidence.
- 05
Answer
Without an AI key, the answer quotes the best-matching sentences verbatim. With a key, a model answers from the retrieved passages only, and every citation is checked to exist in those passages before it is shown.
- 06
Inspect
Selecting a citation scrolls the viewer to the passage and highlights the quoted text.
Security and privacy
- — Uploaded files are processed in memory and are not written to disk or stored on the server.
- — Extracted text is kept only in your browser tab; it is not saved to local storage. Clearing a document removes it and its answers.
- — Document text is treated as untrusted data. It is never executed, and instructions inside a document cannot change how the app behaves; the model has no tools it could be tricked into calling.
Limitations
- — Keyword retrieval can miss answers phrased with different words than the question (no semantic embeddings).
- — The rule-based mode quotes sentences; it does not paraphrase or combine facts.
- — PDF layout (tables, columns) may be flattened during extraction. Scanned PDFs are not supported.
- — Each question re-sends the extracted passages to the server, which bounds document size to keep requests small.