Imagine effortlessly turning human language into the precise commands that power software applications. That's the magic of Natural Language to Code Generation (NL2Code), made possible by the rise of powerful Large Language Models (LLMs). But what happens when we venture beyond common programming languages like Python or C++ and into the realm of Domain Specific Languages (DSLs)? These specialized languages are the backbone of many enterprise applications, allowing developers to write streamlined, targeted code for specific tasks. However, DSLs present unique challenges for LLMs due to their reliance on custom function names, which frequently change and can easily confuse even the most sophisticated AI. This leads to inaccurate code with syntax errors and 'hallucinations'—instances where the LLM invents non-existent functions. In this exploration, we delve into a groundbreaking comparative study that pits two leading approaches against each other: fine-tuning versus optimized Retrieval Augmentation Generation (RAG). Fine-tuning involves training an LLM specifically on a DSL dataset, which yields high accuracy but struggles to keep up with the constant evolution of DSL functions. RAG, on the other hand, dynamically fetches relevant code snippets from a vast database, adapting more readily to new functions. The researchers evaluated both methods using a synthetic dataset mimicking real-world automation tasks involving over 700 APIs. The results reveal a fascinating trade-off: while fine-tuning achieved the best code similarity, RAG excelled in reducing syntax errors, highlighting its potential for handling the dynamic nature of DSLs. This study provides valuable insights into the ongoing quest to bridge the gap between human language and the intricate languages of software, opening doors to more efficient, adaptable, and readily-updated code generation tools.