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.
Last updated:
Short answer: ChatPDF uploads the PDF to its servers and keeps the chat there; SecureRAG parses it with PDF.js in the tab and keeps the page numbers.
What ChatPDF does better
- Nothing to download. Its FAQ states the free plan analyses 2 documents per day and that core features work before you create an account; an account adds saved history and multi-document chats.
- No model fetch on first use. SecureRAG pulls a 23–25 MB embedding model the first time you open it — 120 MB if you switch to multilingual-e5-small for mixed Chinese and English text.
- Fewer per-file restrictions in front of the index. SecureRAG refuses a file over 25 MB outright; hosted services set their own per-plan ceilings, which are worth reading on their site.
- History that follows you. Sign in and the chat exists on another device. SecureRAG has no account, so a new browser profile means importing again.
What happens to a 30-page PDF in each
The ingestion path is the whole story. ChatPDF receives the upload and indexes it server-side, which is why the same chat is available from a phone an hour later. SecureRAG reads the file through the browser File API, parses it with PDF.js so every chunk remembers the page it came from, strips repeated headers and footers, rejoins hard line breaks inside paragraphs, then cuts the text along heading, paragraph and sentence boundaries into roughly 700-character chunks with 15% overlap. Those chunks are embedded in a Web Worker and written to IndexedDB. Ask a question and the query is embedded with the same model, two ranked lists are produced — dense vectors and BM25 keyword matches — fused with reciprocal rank fusion at k=60, thinned by maximal marginal relevance at λ=0.7, and the top 6 chunks are handed to the answering step.
| Question | ChatPDF | SecureRAG |
|---|---|---|
| Getting started | Open the site and upload; no account for core use | Open the site; first load fetches a 23–25 MB embedding model |
| Where the PDF is parsed | ChatPDF's servers | Your browser tab |
| What a citation points at | A page reference back into the uploaded PDF | A chunk with the source page number, plus every sentence carrying its own numbered link |
| Multiple documents | Multi-document chat after sign-in, per its FAQ | Up to 40 files in one library by default, searched together |
| Offline | No | Works offline once cached |
| Ceilings | Per-plan page counts and file sizes; see chatpdf.com | ≤25 MB per file, ≤40 files, ≤200 MB per library, ≤20,000 chunks |
| Unsupported input | Check the vendor's current format list | Encrypted PDFs, scans without a text layer, legacy .doc, and anything over 25 MB are refused |
Page numbers are the part people actually notice
Both tools answer a question about a 30-page contract with a pointer into the document. The difference shows up in the second question. Ask something that requires two clauses in different sections, then click every citation: ChatPDF returns you to the pages it used; SecureRAG returns one citation per sentence, so you can see that the notice period came from page 12 and the exception from page 27, and that the connective sentence between them is a selection rather than a claim the documents make. In "documents only" mode an unsupported question returns "not found in your documents" instead of a plausible paragraph stitched from unrelated pages.
Where SecureRAG loses to ChatPDF
- It costs you a download. 23–25 MB for the default tier, 400 MB–1.0 GB if you enable the generation tier, and that tier runs at roughly 3–8 tokens/s on CPU when WebGPU is unavailable.
- No saved history across devices. Close the browser profile and the library is gone unless you exported it.
- No OCR and no encrypted-PDF support. Password-protected files are out of scope entirely.
- Lower file ceilings than a paid hosted plan: 25 MB per file, 40 files, 200 MB per library, 20,000 chunks, and less again on mobile browsers.
- Prose that loses to a cloud model. A 0.5B–1.5B quantised model answers more tersely and summarises less fluently.
Run the comparison on one of your own PDFs
- Pick a PDF you know well, longer than 20 pages, with a numbered clause and at least one table.
- Ask three questions: one whose answer is a single figure, one that requires two distant sections to be combined, and one the document cannot answer at all.
- Click every citation in both tools. A page reference tells you roughly where the answer lives; a per-sentence citation tells you which sentence is load-bearing.
- Switch the network off and repeat the second question. Only one of the two tools keeps answering.
- Open DevTools → Network, clear the log, and repeat the first question. Compare what each request contains.
Can SecureRAG read a scanned PDF?
No, if the scan has no text layer. Ingestion detects the empty text and refuses the file with an explanation rather than returning an empty answer. OCR is not part of the pipeline; you would need to OCR the file elsewhere and import the text version.
Do I need an account to use SecureRAG?
No account, no tracking pixels, no email. Nothing is stored outside you browser profile, which also means there is no password reset and no support channel that can recover a lost library.
How many PDFs can be searched at once?
Up to 40 files per library by default, 200 MB in total and 20,000 chunks, whichever limit is reached first. Retrieval searches the whole library at once rather than one file at a time, so a question can be answered from two different documents and cite both.
Is the answer quality comparable to ChatPDF?
For extractive answers over clear text, close enough that most people will not notice. For long summaries and fluent rewrites, no — that is the trade you accept in exchange for the file never leaving the device, and the optional generation tier narrows the gap without closing it.