GPT3 model generates a SQL. You can sqldf on top of your data.table. We will be demo'ing at one of the events shortly. BTW, you could do somewhat similar with other LLMs such as GPTJ and GPT NEOX if you have worked with them
They are decently good, I could not find major differences for the cases I was trying. The key is to control the temperature. Make sure it is low, otherwise the randomness increases tremendously. Infact you can feed the same input from openAI into NEOX and it generates results. There are many NEOX open playgrounds that allow you the control the temperature etc.