Pierre Oberholzer

Trying the DDL as you suggest: same issues as in the first trial above. ``` df_information_schema = vn.run_sql(METADATA_QUERY) res = [] for index, row in df_information_schema.iterrows(): col_name = row["column_name"] type...

Good idea. It seems it at least received meaningful info (just showing first 6 columns here).. ```vn.get_training_data()```

Those are dummy column names, the real ones have some semantic meaning. Still, I would expect in any case the generated SQL to be *constrained* on existing column names (even...

Still investigating, but it seems that I am reaching some limit.. 1) Is the entire metadata context passed to OpenAI ? 2) Am I supposed to purge Chroma DB ?...

Thanks - Using ```gpt-4-turbo-preview``` helps. How can I reset the Chroma state, and start again with new context ? Currently I'm accumulating context it seems.

Cool. This helps defining the experiment better. My observation so far is that amongst metadata, ddl and sql, only the latter leads to meaningful generated queries. While this seems promising,...

Sure, if I get to that point.