I am fine tuning GPT on the schema of my database so it can help me create SQL from users natural language questions. I have about 70 fields and some of those fields contain 20-30 lookups. Here is an example user question: How many homes have sold with a foyer?
I was thinking about how I could add some context into the training data around the correct Field and Lookup value to use based on the question. I have an extensive system prompt describing my fields and have a pretty big dataset for fine tuning that includes each lookup value. Has anyone added “context” into the user role? Do you think it is worth adding?
{
“messages”:[
{
“role”:“system”,
“content”:“Please generate an SQL query based on the following user query and provided context.”
},
{
“role”:“user”,
“content”:“How many properties are for sale with an Entrance Foyer?”,
“context”:{
“Field name”:“interiorfeatures”,
“Parameter”:“Entrance Foyer”,
“Parameter description”:“An entry hall or other open area in a house upon entering.”
}
},
{
“role”:“assistant”,
“content”:“SELECT COUNT(*) FROM properties WHERE interiorfeatures LIKE ‘%Entrance Foyer%’;”
}
]
}