This paper examines the application of Bidirectional Long Short-Term Memory (Bi-LSTM) networks in neural source code generation. The research analyses how Bi-LSTMs process sequential data bidirectionally, capturing contextual information from both past and future tokens to generate syntactically correct and semantically coherent code. A comprehensive analysis of model architectures is presented, including embedding mechanisms, network configurations, and output layers. The study details data preparation processes, focusing on tokenization techniques that balance vocabulary size with domain-specific terminology handling. Training methodologies, optimization algorithms, and evaluation metrics are discussed with comparative results across multiple programming languages. Despite promising outcomes, challenges remain in functional correctness and complex code structure generation. Future research directions include attention mechanisms, innovative architectures, and advanced training procedures.

Keywords: code generation, deep learning, recurrent neural networks, transformers, tokenisation