Getting Accurate Results from converting natural language to SQL query using open ai api

I have been writing system role instructions by providing the database schema.
But my schema consists of similar type of column names, say I have 2 columns unserved count and unserved percentage.

When the user prompt is :Identify the top 5 states having largest total of unserved and underserved data
the query generated will be: SELECT unserved count+ underserved count
FROM table name
ORDER BY unserved count+ underserved count DESC
LIMIT 5;

now similarly for the following user prompt: Identify the top 5 states having largest of unserved and underserved data
the query generated will be: SELECT *
FROM table name AS
ORDER BY unserved percentage + underserved percentage DESC
LIMIT 5;

So why does the 2 prompts answer differently I know about the word embeddings that count, total all is related words. How to make those 2 statements get the same response any idea?