Fast path for already-vectorized inputs (eg pyarrow string arrays)

[NickCrews]

I am working with text data in duckdb, doing bulk operations to perform NER on text. I'd like to make these as fast as possible.

In duckdb, you can register python UDFs. You can tell duckdb to call your UDF with either

1. individual vanilla python scalars (eg str in my case)
2. pyarrow arrays, for doing vectorized operations.

Currently, spacy only support as scalar inference API of nlp(str) or a batch API of nlp(Iterable[str]). If I want to use the batch API (which I assume will be faster), then I still have to pay the translation cost of pyarrow array to Iterable[str].

several questions:

1. Since spacy is implemented at the C level, would it be technically possible to add in some short-circuit code that looks for a couple common formats (pyarrow, numpy, pandas arrays come to mind) and avoids the translation cost?
2. Would this actually gain anything? eg perhaps the actual inference stage takes up 95% of the time, so if we improve the ser/deser on either end, we aren't actually going to see much benefit.
3. If both the above look promising, what else should I be thinking about here? Any possibility of this actually happening? What would I need to do to implement this myself/help a core maintainer do it?

Thank you!