Structured output does not follow the specified order of the JSON schema

[mexpex]

Package: lmstudio-1.4.1
LM Studio: 0.3.20(Build 4)

Issue: A structured output does not follow the specified order of the JSON schema provided. Instead the response structure is in alphabetical order. The schema works fine if I use it directly in LM Studio Client.

The current workaround is to add a prefix to the objects in the schema if I want to keep the order.
Tested with various models.

Simple example case:

schema = {
    "type": "object",
    "properties": {
        "name": {"type": "string", "description": "Name of the subject"},
        "age": {"type": "string", "description": "Age"},
        "a_summary": { # a_ to showcasae alphabetical order
            "type": "string",
            "description": "Short summary of the person's greatest achievement",
        },
    },
    "required": ["name", "age", "a_summary"],
}

chat_session = lms.Chat()
chat_session.add_system_prompt("You are a helpful AI assistant")

user_input = chat_session.add_user_message("Tell me about a famous person")

result = model.respond(
    chat_session, config={"temperature": 0.4}, response_format=schema
)

print(json.dumps(result.parsed, indent=2))

In this case "a_summary" ends up first in the response.

[ncoghlan]

JSON schema properties don't have a defined order, as JSON object maps don't have a defined order. The two most common orders across JSON processing libraries are listing order and lexical (sorted) order, but even those aren't guaranteed.

If the property order matters when consuming the data, the most straightforward way of ensuring that happens in an LLM client application is to store an ordered list of known keys and iterate over that instead of directly over the received mapping.

In this case, schema["properties"].keys() would give the property names in listing order, since Python dicts are defined as iterating over their keys in their insertion order.