PDF Gatekeeper

Stage-1 PDF validation, OCR and text extraction pipeline

Checking backend...

How it works

Upload a PDF or provide a public URL. The backend (FastAPI + PyMuPDF) validates the file, analyzes every page, extracts native text, applies OCR to scanned content, and returns normalized page-aware JSON ready for the next processing stage.

GPU & Inference Cost Control

Keep Signal Orbit on CPU. Start paid GPU compute only when needed, then stop it after inference.

Disabled

Reading inference runtime…

Reviewer for this inference session

Off: Ministral extraction + deterministic evidence validation. On: add the independent Gemma reviewer pass. In Serverless mode the reviewer GPU wakes automatically only when the run actually uses it.

Reviewer OFF

Always-on baseline

CPU app + llama.cpp fallback

Automatic protection

Loading cost guard

Hugging Face

HF_TOKEN not configured

RunPod and Hugging Face credentials remain server-side. Stopping a Pod stops GPU compute billing; persistent volume/storage may still have its own charge. Secret tokens are never returned by this API.

Resumable PDF upload

Up to 2 GB · pause/resume · crash recovery · automatic retry

Network online

Initializing recovery

Recovery window

Checking workspace

Checking browser persistence

Initializing resumable uploader and recovery state...

Large-file recovery uses Tus server offsets. Golden Retriever restores upload state after accidental refresh/tab closure when the browser can retain the file handle; after a full browser restart, a recovered “ghost” asks you to re-select the same PDF and then reconnects to the existing Tus upload instead of starting from zero. The 2 GB file body is not copied into browser IndexedDB.

Automatic OCR

OCR only pages that need it after the upload has completed.