Aryan SinghandClaude Opus 5
c6a564ec5d
Every stage now runs client-side: pdf.js renders, a per-pixel diff finds the hidden ink, tesseract.js transcribes it on request, and pdf-lib rebuilds the flattened export. The PDF goes from the file picker into memory and back out as a download, and is never sent anywhere. That makes the privacy claim structural rather than a promise. The server is left with one route — a health check — so there is no code path that can receive a document, nothing is persisted, and no account exists. The OCR model and WASM are staged out of node_modules into public/ at build time so the page never reaches a CDN either. Verified in the browser: a full scan with OCR contacts exactly one origin, its own, and POST /api/upload returns 404 in the built image. Contrast stretching turned out to be the wrong method entirely. Text painted in pure #FFFFFF rasterizes to bytes identical to blank paper, so there is no difference in the image to amplify — and a page with black body text already spans 0-255, so the old server-side normalize() was a no-op. It could never have found the case it advertised. So each page is rendered twice, on white and on black. White ink vanishes into the first and blazes in the second; inverting the black-backed render and taking the darker pixel of each pair puts every mark on the page in view whatever colour it was drawn in. Confirmed against a PDF carrying a genuine white-on-white payload: invisible in "as delivered", flagged in vermilion in "revealed", and transcribed verbatim by OCR. Pages that paint their own opaque background defeat the technique, so that case is detected and reported as unreliable instead of showing a page-sized false positive. OCR is opt-in, as asked: it costs an ~8MB model download and runs at a few seconds a page, so nothing is fetched until the box is ticked. Consequences: - Removed auth, Prisma, Postgres, the upload/render/pixel/ocr/export routes, the storage layer and the signed-in dashboard. - Runtime dependencies drop from 13 to 6, all browser-capable; sharp, @napi-rs/canvas, next-auth, pg and the Prisma stack are all gone. - compose is a single stateless service with no volumes and no config, the entrypoint and its migration step are gone, and the image drops from 868MB to 455MB. - The landing page no longer describes the contrast method it used to claim, since that method provably does not work. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
120 lines
4.8 KiB
Markdown
120 lines
4.8 KiB
Markdown
# XRayPDF
|
||
|
||
**Find the invisible text in a PDF, and strip it out — entirely in your browser.**
|
||
|
||
A PDF can carry text that no human ever sees and every machine reads in full:
|
||
white ink on white paper, half-point type, paragraphs parked outside the crop
|
||
box, a text run hidden behind a full-page image. It renders as nothing. It
|
||
copies, pastes and extracts at full length.
|
||
|
||
XRayPDF surfaces that content, shows you what it found, and hands back a file
|
||
with none of it left. No upload, no account, no server-side processing.
|
||
|
||
## Your file never leaves the tab
|
||
|
||
| | |
|
||
| --- | --- |
|
||
| **No upload endpoint** | The server has exactly one route, a health check. There is no code path that receives a document. |
|
||
| **No database, no disk** | Nothing is persisted. Close the tab and the file, the renders and the OCR output go with it. |
|
||
| **No account** | No sign-in, no email, no session. Nothing identifies you. |
|
||
| **No third-party requests** | The OCR model and WASM are served from the app's own origin, not a CDN. Open the network tab while you scan: every request goes to that origin, and none carry your file. |
|
||
|
||
The honest caveat: the page is still served to you over the network, so you are
|
||
trusting that the code you received is the code described here. If that isn't
|
||
good enough for your threat model, run the image yourself.
|
||
|
||
## How it works
|
||
|
||
**Turning up the contrast does not work.** Text painted in pure `#FFFFFF`
|
||
rasterizes to bytes identical to blank paper — there is no difference in the
|
||
image to amplify, and a page with ordinary black text already spans the full
|
||
0–255 range, so normalizing it is a no-op. Any tool claiming to reveal hidden
|
||
text by boosting contrast cannot find the most common case.
|
||
|
||
So XRayPDF asks the renderer a second question instead:
|
||
|
||
1. **Render each page twice** — at 2× scale on a white backdrop, then on black.
|
||
Same page, same engine, different thing behind the ink.
|
||
2. **Compare them.** White ink vanishes into the first render and blazes in the
|
||
second; ordinary dark ink does the reverse. Invert the black-backed render
|
||
and take the darker pixel of each pair, and every mark on the page comes out
|
||
dark on white, whatever colour it was drawn in. Anything dark there but
|
||
paper-white in the plain render was never meant to be seen.
|
||
3. **Read it back with OCR** *(optional)* — Tesseract over that combined image
|
||
transcribes hidden text alongside the visible copy. Off by default: it costs
|
||
an ~8 MB model download on first use and runs at a few seconds per page.
|
||
4. **Flatten the export** — rebuild the PDF from page images. No text layer, no
|
||
annotations, no object history.
|
||
|
||
One limitation worth stating: a page that paints its own opaque background
|
||
leaves nothing for the black backdrop to show through. XRayPDF detects that
|
||
case and reports the scan as unreliable rather than showing you a page-sized
|
||
false positive.
|
||
|
||
**The trade on export:** a flattened PDF is no longer searchable or selectable.
|
||
Keep the original when you need the text layer; send the clean copy when you
|
||
don't know what is in the file.
|
||
|
||
## Run it
|
||
|
||
### From the published image
|
||
|
||
Multi-arch images (`linux/amd64`, `linux/arm64`) are published to GHCR on every
|
||
push to `main`:
|
||
|
||
```bash
|
||
docker run --rm -p 3000:3000 ghcr.io/geneticglitch1/xraypdf:latest
|
||
```
|
||
|
||
No environment variables, no database, no volumes — the app is stateless.
|
||
Tags: `latest`, `sha-<commit>`, and `v1.2.3` / `v1.2` on release tags.
|
||
|
||
### With compose
|
||
|
||
```bash
|
||
docker compose up -d
|
||
```
|
||
|
||
Set `HOST_PORT` to publish somewhere other than 3000, and `XRAY_IMAGE` to pin a
|
||
tag or digest. To build from source instead of pulling:
|
||
|
||
```bash
|
||
docker compose -f docker-compose.yml -f docker-compose.build.yml up -d --build
|
||
```
|
||
|
||
### Local development
|
||
|
||
```bash
|
||
cd client
|
||
npm install
|
||
npm run dev
|
||
```
|
||
|
||
`predev` and `prebuild` stage the pdf.js worker and the Tesseract core/WASM out
|
||
of `node_modules` into `public/`, so the app serves them itself rather than
|
||
pulling them from a CDN at runtime. Those staged files are gitignored — they are
|
||
build output derived from the lockfile.
|
||
|
||
`GET /api/health` is an unauthenticated liveness endpoint used as the container
|
||
healthcheck. It is the only server route in the app.
|
||
|
||
## Limits
|
||
|
||
- Pages render at 2× scale, capped at 4 megapixels each so large-format
|
||
documents don't exhaust a mobile tab.
|
||
- OCR is English-only; `eng.traineddata` ships with the app.
|
||
- Rendering is `requestAnimationFrame`-driven, so browsers throttle it in
|
||
background tabs. Leave the tab in front while a scan runs.
|
||
- The image is ~455 MB, most of which is the three Tesseract WASM builds
|
||
(~20 MB) plus the language model — a browser downloads only the one variant
|
||
it needs, and only if OCR is switched on.
|
||
|
||
## Stack
|
||
|
||
Next.js (App Router, standalone output) · `pdfjs-dist` · `tesseract.js` ·
|
||
`pdf-lib`. Six runtime dependencies, all of which run in the browser.
|
||
|
||
## License
|
||
|
||
Apache-2.0 — see [LICENSE](LICENSE).
|