docpipe

API Key

Admin Secret (for key management)

Submit documents

PDF files (one or more)

Pipeline mode

Extract categories (default: all — segments classified outside the selection are skipped)

Invoices Delivery notes Orders / offers Contracts Payment docs Everything else

Inference engine (extraction / splitting — OCR always runs on the OCR backend)

OCR model (optional)

Extraction model (optional)

Status

Filter visible (current page)

From

Batch submit from S3

S3 prefix (bulk — enqueues every matching object, chunked server-side; safe to re-run, existing keys are skipped)

S3 keys (one per line)

Pipeline mode

Extract categories (default: all — segments classified outside the selection are skipped)

Invoices Delivery notes Orders / offers Contracts Payment docs Everything else

Inference engine (extraction / splitting)

Published evidence viewers — the interactive report explorer with citation highlighting. Publish one with report_viewer/build.py --publish; it appears here.

Loading…

Inference cost estimator — constants measured on a real 68-page report (OCR 0.176 H100-s/page; extraction 0.088 H100-s/page, 977 in + 116 out tokens/page). Prices as of 2026-07. Serverless cold starts for one-off docs are not modeled (~$0.10–0.30/doc).

Inputs

Documents

Avg pages / doc

OCR backend (self-hosted) Extraction backend

Utilization 75%

Deadline (days)

Estimate

—per document —total

OCR Extraction

—GPU-hours

—GPUs / deadline

—API tokens

Export training data

Stage

Create API key

Webhook tester

Sends a signed sample payload so the receiver (e.g. the label studio ingest) can be verified without running a job.

Launch vLLM model servers on Modal (serverless, scales to zero — dev & demos) or Nebius (preemptible GPU VMs at ~half the hourly rate — batch production). Requires the admin secret.

Launch a server

Set as the active backend for its role once deployed