Using Hermes agent to compare PDF extraction tools
In a firsthand test, was connected to GPT-5.5 and used to read 24 PDFs first, creating a baseline for what the files should say. The files were randomly chosen, only some shared the same format, and many were difficult items such as long thermal-paper invoice printouts. then helped set up a test bench on a DGX Spark to run several tools and against the same files.
Different tools were strong in different ways. Nemotron Parse performed best overall because it could read every file and worked fairly quickly. It still made a few small reading mistakes, as optical character recognition tools often do.
olmOCR stood out for accuracy because it handled common mix-ups such as the letter O and the number 0, but it was quite slow.
Key points
- was connected to GPT-5.5 to read 24 PDFs and create a baseline.
- A DGX Spark was used to test several tools and .
- Nemotron Parse was the strongest overall because it read every file and was fast.
- olmOCR was the most accurate on tricky character mix-ups, but it was slow.
- A useful pattern is to make a baseline first, then test other tools against it.