edgedb
Add more pgvector types.
Add support for sparsevec and halfvec.
halfvec is handled as a float array (converted from float32 to float16 or vice versa as needed).
sparsevec is handled as a dict with non-zero values and "dimensions" specified in it.
The halfvec codecs need more robust conversion for float16. This seems to outline a good algo that I can implement here: http://www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf. As a bonus, I don't even need to handle NaN or Inf values because they are not valid in vectors (I checked with the actual pgvector types). So it's a little bit fewer corner cases to cover.
Sometimes _Float16 might be available via #include <float.h>, which should take care of casting between float/double and float16, but it would still need a fall-back.