I am working on a POC to convert Natural language to SQL. I have used phi3 and now planning to use sqlcoder as part of the llm. All this are set up via ollama which I am running on docker.
The one thing I stumble upon is how to train my custom data ( Database Schema).

1. I have used the System Prompt by providing the necessary instructions and response, it does work.but I believe it has very limited token size, which will not suffice for my usecase.

2. Used phi3 model’s instructions to create a custom model by just changing the instructions, but the output latency very high.

Any pointers for on how to feed/fine-tune the custom schema of 500+ tables to LLM, would be helpful.
As I can use Sqlcoder which is already trained on text-to-sql, do I still need to train it on my custom schema?