I was thinking of something similar — instead of just two passes, couldn’t you also store different quantized values? If you have thousands of documents, you could narrow it down to a handful with a few bit-wise Hamming comparisons before doing a full cosine similarity on just the rest. If you hand more than one bitmap stored, you’d have fewer comparisons at each step too.

Would this work?

Ooh that does look nice. However, that license is a deal breaker for me.