Automate common tasks with the help of AI
- Name scanned PDFs
- Add meta data to scanned PDFs
- Convert images to PDFs
- Join Images to one PDF
- Join PDFs to one PDF
- Java 21
- https://github.com/ocrmypdf/OCRmyPDF
- https://ollama.com/
- docker e.g. https://docs.docker.com/desktop/
- Try qwen3:4b or gemma3:4b, if you have not many resources
- If you have plenty of RAM 24GB / VRAM 16 GB try first granite3.3:8b next qwen3:8b
- If you have more RAM (>=32GB) you also test gemma3:12b if it creates better results
| LLM | Rechnung Hotel Stern | LIDL Rechnung |
|---|---|---|
| qwen3:4b | 15/40 = 37% | 8/32 = 25% |
| qwen3:8b | 15/40 = 37% | 8/32 = 25% |
| qwen3:14b | failed | failed |
| gemma3:4b | 13/40 = 32% | 8/32 = 25% |
| gemma3:12b | 12/40 = 30% | 11/32 = 34% |
| mistral:7b | 12/40 = 30% | 9/32 = 28% |
| granite3.3:8b | 15/40 = 37% | 12/32 = 37% |
| llama3.1:8b | 12/40 = 30% | 9/32 = 28% |
| deepseek-r1:14b | 13/40 = 32% | failed |
| gpt-oss:20b | failed | failed |
=> deepseek-r1:14b is very unstable works only sometimes
Results differ a bit from run to run
docker run -i --rm jbarlow83/ocrmypdf --skip-text -l deu - - < in.pdf > out.pdf
ProcessBuilder pb = new ProcessBuilder(
"docker", "run", "-i", "--rm", "jbarlow83/ocrmypdf", "-l", "deu", "-", "-"
);
for filename in ./*.pdf; do
docker run — rm -v=$(pwd):/tmp jbarlow83/ocrmypdf -l deu /tmp/${filename#./} /tmp/out/${filename#./}
done