NL to SQL Architecture Alternatives

Hello YaoJing - two possible ways you could handle this situation:
1. Embed all table schemas, as well as the table name and table description (if available) into a vector DB. Include one entry for each table. Then at run time, query the vector DB based on the user's question and include the top n records as context. 

2. When the user asks a question, query a cheap, fast model, such as GPT 3.5 T or Phi 3 and include all table names and description. Ask the LLM to select the top n tables which are most relevant to the user's question. Then include these tables in the subsequent request to generate the SQL query.

Let me know if this works for you.