We played with Langchain before deciding on your solution. I think the biggest value add from Dot here, aside from the prompt engineering, is the infra around s

We played with Langchain before deciding on your solution. I think the biggest value add from Dot here, aside from the prompt engineering, is the infra around spinning up the meta tables that describe columns. I think you could get good results with Langchain if you precomputed them like Dot does. Having metadata describing the content (which Langchain tries to do on the fly by default by querying the top 5 or so rows, which is slower and leads to… mixed results) turns the tool from basically useless to pretty good instantly.