Gen AI with structured output - `propertyOrdering` in Python vs. Java implementations

[j-nc]

Hi, when using gen AI with schema-constrained output, I've noticed in google/genai/_transformers.py, there's explicit handling of propertyOrdering (aka. property_ordering), while it seems that this is not the case in the Java implementation.

In the Python implementation the logic for including it is fairly simple: if propertyOrdering is missing, add it:

      if (
          len(properties.items()) > 1
          and order_properties  # True by default
          and 'propertyOrdering' not in schema
      ):
        schema['property_ordering'] = list(properties.keys())

I wouldn't bring this up if it didn't actually result in AI returning different, and better responses with it, vs. without it. Specifically, in my Java app, with the exact same prompt and basic schema (without explicitly setting propertyOrdering), AI consistently returns fewer fields than it does when invoked via the Python app.

Now I've noticed this field is labeled as non-standard, at least in v1.51.0, which I have locally:

    /**
     * Setter for propertyOrdering.
     *
     * <p>propertyOrdering: Optional. Order of properties displayed or used where order matters.
     * This is not a standard field in OpenAPI specification, but can be used to control the order
     * of properties.
     */

Despite that, I've now added propertyOrdering to my schema (in Java). (Note, I have Vertex AI set to true in both cases, if that makes a difference.)

Is there any general information on this? In particular, is propertyOrdering expected to have an effect on structured output generation in Vertex AI, and should the Java implementation also be adding it automatically when it isn't explicitly provided?

Thanks.

[hemasekhar-p]

Hi @j-nc, thank you for bringing this to our attention. propertyOrdering is used during structured output generation to ensure deterministic field sequencing. To achieve cross SDK feature parity with Python, opened a PR to automatically populate propertyOrdering with the keys of properties whenever a schema contains multiple fields and no custom order was specified. If an explicit propertyOrdering is provided during schema construction, your custom field ordering will continue to be strictly preserved.

In the meantime, you can manually set propertyOrdering when building your Schema object as a work around until the fix is released.