Thanks for sharing this! The context-based approach is interesting. Maintaining state from past successful parses to resolve ambiguity is essentially what we do

Thanks for sharing this! The context-based approach is interesting. Maintaining state from past successful parses to resolve ambiguity is essentially what we do too, though we batch it (scan first 50 values) rather than doing it incrementally. The incremental approach has the advantage of adapting mid-column if formats shift, which we've seen in datasets that were manually concatenated from different sources. Will take a look at the repo.