Notes from Discord discussion:
There are some interesting quirks due to the nature of LanceDB; you need to specify a schema for your metadata so that it can be turned into a pyarrow struct (and thus be searchable). Also the precision of the embedding vectors seems to be limited to an error of around 10^-7 compared to it's python form - feels like some sort of difference between the float representations.
Please take a look when you have time, and let me know if I should change anything.
I've updated the pip install block to be tagged as shell, got in touch with the LanceDB team and sourced the official 'light on dark' logo, and added a link to the new API docs.
This PR adds the LanceDB Haystack Document Store from https://github.com/alanmeeson/lancedb-haystack.
Notes from Discord discussion: There are some interesting quirks due to the nature of LanceDB; you need to specify a schema for your metadata so that it can be turned into a pyarrow struct (and thus be searchable). Also the precision of the embedding vectors seems to be limited to an error of around 10^-7 compared to it's python form - feels like some sort of difference between the float representations.
Please take a look when you have time, and let me know if I should change anything.
Thanks!