Deep Dive into JSON Data Formats, Serialization & AI Token Optimization

How Modern Large Language Models Process JSON

Large Language Models (LLMs) do not read text character-by-character or as raw binary trees. Instead, text is parsed into sub-word chunks called Tokens using algorithms like Byte-Pair Encoding (BPE). Because developers frequently pass large structured JSON datasets as context to AI models, inefficient formatting drastically increases API costs and exhausts model context windows.

Techniques for Optimizing JSON for AI Context Windows

• Structural Minification: Eliminating unnecessary whitespace and line breaks saves between 15% and 30% of total tokens without losing data fidelity.
• Array-of-Arrays Compression (Tabular JSON): When transmitting uniform lists of objects, repeating property keys on every single object creates massive token bloat. Converting objects to a schema header and value arrays can reduce token usage by up to 50%:

  // Traditional Bloated JSON (120 tokens)
  [
    {"id": 1, "name": "Alice", "role": "Engineer", "dept": "Core"},
    {"id": 2, "name": "Bob", "role": "Designer", "dept": "UI"}
  ]
  
  // AI-Optimized Tabular JSON (60 tokens - 50% savings)
  {
    "cols": ["id", "name", "role", "dept"],
    "rows": [
      [1, "Alice", "Engineer", "Core"],
      [2, "Bob", "Designer", "UI"]
    ]
  }