Tokenization is the first step in preparing structured data for embedding, especially when dealing with formats like JSON. Traditional tokenization methods, such as Byte-Pair Encoding (BPE) or WordPiece, break down text into smaller units. However, they struggle with the non-alphanumeric characters in JSON. For example, a JSON snippet like `"usd": 10` turns into fragmented tokens, which creates a low signal-to-noise ratio. In natural language, most words carry meaning, but in structured data, many tokens represent structural elements that don't add semantic value. This leads to embeddings that fail to capture the true relationships in the data. The attention mechanism in Transformers highlights this issue. While it can effectively link relevant tokens in natural language, it struggles with the structural syntax of JSON. As a result, the semantic intent of the data becomes obscured. Mean Pooling, the process of averaging token vectors to create a single embedding, can further muddle the representation if a significant portion of the tokens are structural. For instance, if 25% of tokens in a document are just punctuation, the final vector is skewed away from its true meaning. Flattening JSON data into a more natural language format enhances the embedding process. By converting structured data into sentences, like describing a product with its attributes, the number of tokens decreases and the semantic clarity improves. This article presents a method for transforming JSON attributes into a coherent product description. A test using the all-MiniLM-L6-v2 embedding model demonstrated that this flattening step can improve retrieval metrics significantly. In an experiment with the Amazon ESCI dataset and 5,000 queries, flattening the structured data boosted recall and precision by about 20%. This emphasizes the importance of effective data preparation for optimizing semantic retrieval systems.