I'm not so sure if vector similarity is the right way to think about as opposed to classification but I do think we need an AI "grep"
This could be zero shot "show me paragraphs about unsuccessful construction projects" or one shot or many shot (e.g. have an -option to specify a training file or a model file) It should have some understanding of files in tabular or near-tabular formats such as CSV and JSON and probably be joined to an engine similar to awk or jq so it can effectively answer "When was the temperature greater than 20 degrees centigrade?" and it figures out which field is the temperature field and maybe even if the units are centigrade or something else.
Note I've had very good luck doing text classification using the kind of model people use for vector similarity (e.g. https://sbert.net/) but piping the results through a classical ML model such as the SVM,
I'm not so sure if vector similarity is the right way to think about as opposed to classification but I do think we need an AI "grep"
This could be zero shot "show me paragraphs about unsuccessful construction projects" or one shot or many shot (e.g. have an -option to specify a training file or a model file) It should have some understanding of files in tabular or near-tabular formats such as CSV and JSON and probably be joined to an engine similar to awk or jq so it can effectively answer "When was the temperature greater than 20 degrees centigrade?" and it figures out which field is the temperature field and maybe even if the units are centigrade or something else.
Note I've had very good luck doing text classification using the kind of model people use for vector similarity (e.g. https://sbert.net/) but piping the results through a classical ML model such as the SVM,
Thanks! That’s a really interesting perspective. I hadn’t thought about the classification angle before.