Problem

Extracting structured data from OCR text is noisy and layout-sensitive.

Approach

  • Use layout-preserved text
  • Prompt engineering with examples
  • Post-processing using bounding boxes

Key Learnings

  • Prompt quality matters more than model size
  • Layout context improves accuracy significantly

Conclusion

Hybrid systems (LLM + rules) work best in production.