[otagekki]

Impressive demonstration. Had I had the same set of tools just a few years back, maybe I'd have kept my old job a little longer.

Jokes aside, there is only so much complexity that can be held. Try writing a financial regulatory report from the data present in your enterprise ERP (whose schema you'll eventually have to feed prior or in addition to your prompt).

Granted you can also do it in small steps using only subsets of the whole schema, but that still requires non-trivial oversight by a human expert. Plus, user needs are not set in stone and are not exempt from contradictions. The human might challenge those contradictions, the machine will not.

Infinite-length context could address prompt size related issues, but hallucinations on imaginary columns and flat-out wrong calculations, which cannot be avoided by humans either, will happen past a certain point.

Anyway I am not sure any sensible CFO or Financial department manager would give results produced by an unauditable automated system that takes in English and spits out numbers on a spreadsheet: if errors are missed or bugs arise, who's responsible? (the AI _is_ the expert, remember)

[vishalpatil3435]

The human in the middle is the key part of our solution. The problem with SQL is it takes only 5 minutes to think of the query logic and 25 minutes to force fit it into SQL syntax. The tool focuses more on the latter.