Skip to main content

PDF OCR pipeline simplification — faster, cheaper, same API

· 2 min read
Chao Cheng
Chao Cheng
Software Engineer @ Cobbling AI

We just shipped a backend overhaul of PDF OCR. The user-facing API is unchanged — your existing integrations keep working — but everything behind it is leaner.

What changed (and what didn't)​

Same

  • Endpoints, authentication, request and response shapes.
  • Webhook events and Svix signature verification.
  • Pricing and subscription tiers.

Faster behind the scenes

  • Pages now go straight from PDF to OCR. We dropped the intermediate PDF→PNG conversion that used to run in AWS Lambda for every page.
  • Worker count for the OCR pipeline went from five to two, which removes hops, queues, and retry surface area.
  • Per-page latency in our internal benchmarks improved noticeably, and tail latency on bursty workloads is much tighter now that fewer queues need to drain.

Why we did it​

The old pipeline was three-staged: split the PDF into per-page PDFs, convert each page to PNG (Lambda), then OCR the PNG. The PNG conversion cost real money and added a queue retry hop where, on a transient failure, we'd redo a step that didn't need redoing.

The OCR model already accepts PDF inputs natively. Once we wired the page-OCR worker to send page PDFs directly, the PNG conversion step became dead weight. Removing it eliminated:

  • One Lambda function and one API Gateway.
  • Two Cloudflare Queues (and their dead-letter queues).
  • Three of the five OCR-pipeline workers.

Do I need to change anything?​

No. Same endpoints, same auth, same payloads.

If you want a slightly simpler upload path, we now also document the single-request multipart upload at https://api.commapdf.cobbling.ai/v1/parsing/upload alongside the existing presigned-URL flow at https://api.pdf-ocr.cobbling.ai/v1/upload/direct. Both still work; pick whichever fits your client.

See the refreshed PDF OCR documentation for the full reference.

As always, ping us if anything looks off — we read every report.