3. Existing Code Generation Tools

Auto code generation is an active research area though its success is still limited compared to natural language text generation. Unlike natural language, code has a strict logical specification and a limited syntactical space. But despite the reduced complexity, even a minor mistake can drastically change the meaning of the code or make it completely invalid[Espejel et al., 2024]. In an attempt to achieve better results, there have been numerous publications on using LLM+RAG for code generation since RAG has gained in popularity as a lighter and simpler alternative[Hostnik and Robnik-Šikonja, 2024]to fine-tuning[Bassamzadeh and Methani, 2024]. Most such research relies on a relatively large corpus of RAG input data. One study includes not only code samples, but also API documentation and Stack Overflow articles[Shin et al., 2024] while another has incorporated web search results and execution feedback[Su et al., 2024]. In addition, many have resorted to manual input data manipulations, including tagging[Li et al., 2025], and manual filtering[Wang et al., 2025], in order to improve the accuracy of generated code.