I built this library because I wanted a way to clean messy text data and extract PII without sending sensitive information to cloud APIs or dealing with brittle regex patterns.

The tool runs quantized models locally via llama.cpp and uses GBNF grammars generated from Pydantic models. This forces the LLM to output valid JSON strictly adhering to the schema, which solves the reliability issues common with small models. It currently supports Pandas and Polars dataframes and works with any GGUF model.

It is still an early alpha, so performance on older CPUs might be a bottleneck compared to standard string manipulation, but I found it useful for semantic extraction tasks where regex fails. I would appreciate any feedback on the implementation or suggestions for optimization.