Skip to content

One GPU · Real receipts · Saved evidence

Local document parsing, with measured results

Run NVIDIA Nemotron-3.5-Parse on receipt images, save its output, and check the extracted text and numbers.

This small experiment uses 24 CORD receipts and one RTX 4090. How fast? About 2 receipts per second when sent one at a time, or 6 per second with four in progress at once. That is almost three times as many receipts processed in the same time. The typical receipt took about 0.4 seconds on its own and 0.6 seconds when sharing the GPU. These are warmed-up test results; startup is excluded, and fast output can still contain wrong amounts.

Start with the saved results, or run the parser on your own GPU.

Inspect saved results Run on your GPU

What runs locally

  1. STEP 1Receipt images

    Download the fixed 24-image corpus. Verify its annotations and hashes.

  2. STEP 2Nemotron Parse

    Send each image to the parser, served locally through vLLM.

  3. STEP 3Saved outputs

    Keep raw text, completion status and request timing.

  4. STEP 4Compare text and numbers

    Score saved output against the annotated reference. This step needs no GPU.

Optional second-model review: Judgevet → Typevet → Gemma can review the saved text and reference. Its judgments are advisory and stay separate from measured scores.

Nemotron Parse returns text, layout markers and table structure. This experiment evaluates extracted text and numeric strings, not layout or table-structure accuracy.

What the experiment found

All 48 measured requests completed. One amount still changed.

The receipt says 79,500. The parser returned 79,300. Request completion does not establish correct extraction.

Concurrent requests Completed Throughput Median latency Exact numeric matches
1 24/24 2.12 pages/s 415 ms 212/241 (88.0%)
4 24/24 5.99 pages/s 634 ms 212/241 (88.0%)

Throughput means how many receipts finish each second across all requests. Median latency is the middle receipt's response time: half finished faster and half slower. Handling four at once increased the total processing rate, but each receipt typically waited a little longer.

These are two short sweeps over the same 24 unique receipts on one RTX 4090. The image cache was warmed. Throughput includes scoring and evidence writes. More concurrency increased throughput and median latency; numeric recall stayed the same.

CORD annotates selected content, not every visible word. Numeric presence does not establish correct field association. These observations do not establish general financial accuracy or production capacity.

Read the measured results and failure cases · Understand the scoring rules

Choose your next step