Pdf to JSON, 3 months in.
This covers generation capability or on-device inference progress — worth tracking for model efficiency, deployment cost, and application openings.
A user is seeking solutions for accurately converting PDF data to JSON format after three months of development.…
Hello all, it has been 3 months since I made the initial post, where I wanted ideas to try out. The subreddit has been amazing with responses, and the most success I had was using pymupdf4llm or Docling. I have been sticking to docling for how accurate it is, but I’ve been stuck for a while now.
After docling parses the document, creates a markdown of everything it finds, I have an LLM (Qwen 3.5-9B, Q8, CTX 64k) to having map the data into JSON (the particular json format I need it in, universal for all files I parse). The issue is that, although the markdown contains all the data needed, the mapping is not always correct, like counts, amounts etc are often times wrongly labelled, or not understood from the files, it often replaces values it should not have.
How do I address the last bit? I tried prompt engineering but whenever i change the prompt, I have a new problem, the changes are no way generalized.
I slowly am beginning to realise this might be an LLM capabilities issue, rather than the workflow issue. For this to work I sort of need local models and can’t really rely on APIs, and this LLM configuration is the best I could run on my current system.