Use of reranker and Question about embeddings of DDL, SQL matched against Query

[sourabh-patil]

Hi,
Thanks for sharing this work. I had a question, which might be naive, I did not get how we get expected embeddings of DDL or SQL if we match it against natural language Query? Can anyone please shed some light on it? As per the base class of vanna, we first get the similar embeddings matched against stored (or here referred as trained) SQLs, DDLs and documentation.

[zainhoda]

Hi @sourabh-patil -- would you mind adding some additional context to your question?

Have you seen this diagram?

[sourabh-patil]

Yes, I have gone through this diagram. The question was how we are getting expected embeddings for DDL (which is not a natural language) that are similar to the Question (which is asked in natural language). From this diagram, in train phase we are storing DDL in vector db and for SQL we are storing it as question and SQL pair. At the time of Ask, we fetch embeddings similar to asked question from all three collections DDL, SQL and documentation. So, how do we make sure that DDLs which are stored (which are not in natural language) should match exactly with the question asked (which is in natural language). So do we assume that some words from question will match with some words in DDL? If yes then are there any better embedding models for this part of retreiver?

In case of SQL, we are storing question and sql as pair so we will probably get back the expected SQL.

P.S. Should we use a reranker after we fetch the data from DB? I feel it should improve performance atleast in case of Documentation and SQL