I have a system that currently processes natural language queries and generates insights in charts and narratives formats using Python code rules. However, as the number of question types increases, the system and rules are becoming more complex, and changes in the question structure often lead to failures.
I also attempted using a Natural Language to SQL approach by refining the T5 model with domain-specific NL questions and corresponding SQL queries. However, this approach did not perform well, especially for complicated questions like diagnostics and forecasts.
After completing the first week course, I started wondering if it would be feasible to create a domain-specific Language Model (LLM) to address these challenges. There are two potential approaches:
- Developing a domain-specific LLM trained on NL to SQL or based on question types, which could potentially improve the system’s performance and adaptability to different queries.
- Alternatively, training an LLM to directly generate the current JSON format, which is currently being created by complex rules, thereby simplifying the system and making it more efficient.
Given the complexity of the problem, I’m uncertain about the feasibility of these approaches. Before diving in, I wanted to seek suggestions and guidance on how to proceed. Any insights or advice would be greatly appreciated. Thank you.