Latency Overview
Cardinal is optimized for speed and scale. Each page typically takes around 15 seconds to process with our base model, regardless of size or complexity. Because all pages are processed in parallel with no rate limits, total latency depends primarily on the slowest page rather than the total number of pages.Typical Latency
| Document Size | Expected Latency |
|---|---|
| 1 page | ~15 sec |
| 10 pages | ~15 sec |
| 100 pages | ~15 sec |
| 1000 pages | ~15 sec |
Notes
- Pages run fully in parallel, so overall latency does not scale linearly with page count.
- Base model processing takes approximately 15 seconds per page for standard OCR.
- Additional features can increase processing time:
- Barcode detection may add several seconds per page
- Post-processing and enrichment typically add ~10–20 seconds on top of base latency
- Advanced passes (dense tables, handwriting, annotations) may add ~10–25 seconds per page
- Schema alignment and bounding boxes can add additional processing time
- There are no rate limits — you can process as many pages as you want simultaneously.
- Skew correction and other advanced passes may add a few seconds of processing time per page.
These are approximate numbers. For production workloads, we recommend uploading a sample batch to benchmark your own latency.