Generating structured output

Working with Llama 3

Imtihan Ahmed

Machine Learning Engineer

Structured output in JSON

Example: Llama responses might be input to a dashboard
Plain text responses won't work ❌
We need structured outputs ✅

Employee of food and beverage analytics company

Diagram of automating reports for restaurants

JSON responses with chat completion

response_format = {"type": "json_object"}


message_list = [
       {"role": "system", # System role defined as market analyst
        "content": "You are a food industry market analyst. You 
         analyze sales data and generate structured JSON reports 
         of top-selling beverages."},

       {"role": "user", # User role to pass the request
        "content": "Provide a structured report on the top-selling beverages 
         this year."}
        ]

JSON responses with chat completion

output = llm.create_chat_completion(
         messages = message_list,

         response_format = "json_object"

)

Response format specified as JSON
Llama generates structured response, no free-flowing text

Extracting the JSON response

print(output['choices'][0]['message']['content'])

{
  "report_name": "Top-Selling Beverages 2024",
  "top_beverages": [
    {
      "rank": 1,
      "beverage_name": "Coca-Cola Classic",
      "sales_volume": 2.1,
      "growth_rate": 1.9
    },... ]
}

Defining a schema

response_format = {
"type": "json_object",

"schema": {

    "type": "object",

    "properties": {
           "Product Name": {"type": "string"},
           "Category": {"type": "string"},
           "Sales Growth": {"type": "float"}}

           }

}

Can specify a schema: rules to define how the data should be formatted

Defining a schema

output = llm.create_chat_completion(
         messages = message_list,
         response_format = response_format)  

print(output['choices'][0]['message']['content'])

{
  "Product Name": "Coca-Cola",
  "Category": "Soft Drink",
  "Sales Growth": 12.5
}

Let's practice!

Working with Llama 3