Three ways to ask a document a question, and where your file ends up in each
NotebookLM, ChatPDF and ChatGPT all answer questions about a document you hand them, and all three copy that document onto a company server before anything is indexed. SecureRAG does the same job inside one browser tab: bge-small-zh-v1.5 (~25 MB) or all-MiniLM-L6-v2 (~23 MB) embeds your text locally, retrieval fuses vector search with BM25, and the only outbound request the application ever makes is the model download. These three pages compare each hosted tool against that approach, including the cases where the hosted tool wins.
Last updated:
NotebookLM keeps your sources in a Google account. SecureRAG keeps them in the tab.
NotebookLM is a hosted notebook: you add sources, it indexes them on Google's servers and answers with citations that open the source passage. SecureRAG keeps the retrieval half of that workflow and moves it into the browser — the same file is parsed, chunked and embedded locally with a 23–25 MB model, and the library never leaves the machine. This page compares the two on storage, citations, offline behaviour, limits and writing quality, and names the four things NotebookLM does better.
ChatPDF answers from an uploaded PDF. SecureRAG answers from the same PDF without moving it.
ChatPDF answers questions about a PDF you upload, with page references back into the document, and its own FAQ states that you can start without an account. SecureRAG covers the same ground — PDF parsing with page numbers preserved, hybrid retrieval, per-sentence citations — with the file staying in the browser tab. This page compares the two on setup, page anchoring, document ceilings and failure cases, and is explicit about where the cloud tool is the more practical choice.
ChatGPT reads an uploaded file on OpenAI's servers. Here is what that changes.
Upload a PDF to ChatGPT and it is parsed, chunked and searched on OpenAI's infrastructure, with the answer written by a frontier model. SecureRAG does the parsing, chunking and searching in the browser tab and answers from your own retrieval, with the file never leaving the device. The interesting comparison is not raw capability — ChatGPT wins that outright — but what happens to the document, which setting decides it, and which jobs only a hosted model can do.
The dividing line is not features. It is the copy of the file, and who holds it after you close the tab.
How these comparisons were written
Every claim about a competitor comes from that vendor's own documentation, read in September 2026: Google's NotebookLM help pages, the chatpdf.com FAQ, and OpenAI's help centre. Where a number changes with the plan — daily quotas, page ceilings, per-file size limits — this page names the pricing model and sends you to the vendor for the figure instead of quoting something we cannot keep current. Claims about SecureRAG are quoted exactly as the tool ships, limits included.
| Question | NotebookLM | ChatPDF | ChatGPT file upload | SecureRAG |
|---|---|---|---|---|
| Where the file is parsed | Google's servers | ChatPDF's servers | OpenAI's servers | Your browser tab, in memory |
| Account needed | Google Account, signed in | No account to start; sign-in for saved history and multi-document chats | OpenAI account | No account, no tracking pixels |
| Works with the network off | No | No | No | Yes, once the model is cached |
| What a citation points at | A numbered inline citation that opens the source passage | A page reference back into the PDF | Quoted passages inside the reply; no page anchor for uploaded files | A numbered link per sentence to the chunk; PDF chunks keep page numbers, DOCX chunks keep heading levels |
| What writes the answer | A large cloud model | A large cloud model | A large cloud model | Extractive selection by default; optional Qwen2.5 0.5B–1.5B at 4-bit, 400 MB–1.0 GB, roughly 3–8 tokens/s on CPU |
| File ceilings | Per-plan source and notebook limits; see Google's help pages | Per-plan page and size limits; see chatpdf.com | Per-plan size and token limits; see OpenAI's help centre | ≤25 MB per file, ≤40 files, ≤200 MB per library, ≤20,000 chunks per library |
The copy is the part people forget to price in
Upload a 40-page severance agreement to a hosted tool and you now have two documents: the original on your disk, and a copy on someone else's. That copy lands in a storage bucket, gets picked up by a backup rotation, and is answerable to whatever legal process the vendor is subject to. Deleting the chat later removes the visible entry, not necessarily the object in the backup set. SecureRAG reads the file through the browser File API into an ArrayBuffer, splits it at roughly 700 characters with 15% overlap along heading, paragraph and sentence boundaries, embeds each chunk in a Web Worker and writes vectors to IndexedDB inside your browser profile. Clearing site data removes the whole library, and there is no second copy anywhere to chase.
When the local option is the wrong choice
- You want an audio overview, a mind map or a generated briefing document. Hosted notebooks ship these; SecureRAG does not.
- You need the same library on a laptop and a phone without re-importing. SecureRAG has no account, so there is nothing to sync.
- Your documents are scans without a text layer. The ingestion step detects the missing text and refuses the file; OCR is not in the pipeline.
- You need the most fluent prose available. A 0.5B–1.5B quantised model writes plainly next to a frontier cloud model, and the first load costs 400 MB–1.0 GB.
- You need to hand a colleague a link. A static page with no backend has no share link to hand over.
Choose by the question you are actually asking
- If the answer has to be quotable later — a contract clause, a policy section, a filing paragraph — start with the tool that never moves the file: open the tool, drop two PDFs, and check that a citation marker lands on the right paragraph before you rely on it.
- If you need to share a source with a team, or listen to a summary on the way home, a hosted notebook fits better. Use it deliberately: redact first, or upload only documents you would be comfortable attaching to an email.
- If you already pay for a cloud assistant and merely need a quick read of one file, keep using it. You are trading a copy on a server for convenience, and that is sometimes the right trade.
- If the file is covered by an NDA, a works-council agreement or a client contract with a confidentiality clause, run SecureRAG with strictness set to "documents only". Ask something the text cannot answer and confirm you get "not found in your documents" rather than a plausible guess.
Does SecureRAG make no network requests at all?
It makes one. On first use it issues a GET for the embedding model weights from a public model host (HuggingFace or a mirror); the request carries the model file name and nothing else — no document, no question, no identifier. After that it works offline once cached, and switching DevTools to Offline will not stop it.
Why not use a cloud tool and delete the file afterwards?
Because deletion runs on someone else's schedule. Retention windows and backup rotation are policy decisions you cannot inspect from the outside. A file that never leaves the device has no deletion window to track.
Are the competitor facts on these pages verified?
They are taken from each vendor's own published documentation as of September 2026: Google's NotebookLM help pages, the chatpdf.com FAQ and OpenAI's help centre. Where a figure varies by plan, the page defers to the vendor rather than quoting a number that will be wrong next quarter.
Is SecureRAG a drop-in replacement for these tools?
No. It replaces retrieval and citation over documents you own: local parsing, structure-aware chunking, hybrid search and sentence-level citations. It does not replace audio overviews, Drive source syncing, notebook sharing, image understanding or frontier-model prose.