fix: only set additionalProperties on object schemas

[vidigoat]

What this fixes

set_additional_properties_false_json_schema (used by the OpenAI backend to make a schema strict-compatible) sometimes injects "additionalProperties": false into non-object schemas. The OpenAI structured-output API rejects additionalProperties on anything that isn't an object, so a legitimate schema fails with a 400 invalid_request_error.

Reproduction

from typing import Literal
from pydantic import BaseModel
from outlines.models.openai import OpenAITypeAdapter

class Node(BaseModel):
    name: str
    kind: Literal["object"]   # pydantic emits {"type": "string", "const": "object"}

schema = OpenAITypeAdapter().format_output_type(Node)
kind = schema["response_format"]["json_schema"]["schema"]["properties"]["kind"]
print(kind)
# -> {'const': 'object', 'title': 'Kind', 'type': 'string', 'additionalProperties': False}
#                                                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ invalid on a string field

Any scalar value equal to the string "object" triggers it — a Literal["object"] field (emitted as const), or a field whose default / title / description is "object".

Root cause

jsonpath_expr = jsonpath_ng.parse('$..*')
for match in jsonpath_expr.find(schema):
    if match.value == 'object':                       # matches ANY value == "object"
        if 'additionalProperties' not in match.context.value:
            match.context.value['additionalProperties'] = False  # mutates the parent

It keys off the string value "object" appearing anywhere in the schema and then mutates that node's parent, instead of keying off the JSON Schema keyword {"type": "object"}.

Fix

Recurse over the schema and set additionalProperties only on dicts whose type is "object". Nested object schemas are still handled; already-present additionalProperties values are preserved. This also drops the jsonpath_ng usage in this helper in favour of a plain recursive walk.

Test evidence

Added two regression tests in tests/models/test_utils.py:

• test_set_additional_properties_false_json_schema_string_value_object — fails on main, passes with the fix (a string field with const/title "object" must not get additionalProperties).
• test_set_additional_properties_false_json_schema_nested_objects — documents that nested object schemas still get additionalProperties.

$ pytest tests/models/test_utils.py tests/models/test_openai_type_adapter.py -q
16 passed

ruff check and mypy --allow-redefinition both pass on the changed files.

[RobinPicard]

Thanks for opening a PR! There's a conflict to solve with the main main as the file you are modifying has been updated. Let's remove the jsonpath_ng dependency too as you are removing the only use of this library.

[Sanjays2402]

This misses nullable object schemas whose type is ["object", "null"]. Those nested objects keep no additionalProperties, so the strict OpenAI/Mistral schema is rejected; the equivalent check needs to handle lists containing "object" too.

[vidigoat]

Thanks @RobinPicard! Rebased onto main so the conflict is resolved, and removed the jsonpath_ng dependency (from both pyproject.toml and uv.lock) since this was its only use — uv lock --check passes.

@Sanjays2402 good catch on nullable objects — the check now also matches type lists containing "object" (e.g. ["object", "null"]), so nested nullable object schemas get additionalProperties: false too. Added a regression test for that case alongside the existing ones.

[ErenAta16]

Reproduced the old bug directly before checking the fix, since the whole thing hinges on whether the old jsonpath walk really did hit non-object schemas.

It does. Running the old $..* version against the PR's own repro ({"type": "string", "const": "object", "title": "object"} nested under an object) adds additionalProperties: False straight onto the string schema, because $..* visits every value in the tree, "object" shows up as the value of const and title, and match.context.value is then the string schema itself. So the fix is addressing a real 400 from the OpenAI side, not a theoretical one.

The rewritten _walk keys off node.get("type") instead of any value that happens to equal "object", which is the correct signal. Checked it against a few cases beyond the one in the test:

• The repro's string-with-const: "object" field: correctly left untouched.
• Plain nested object: still gets additionalProperties: False.
• Nullable object with "type": ["object", "null"]: still gets it, the list branch is preserved.
• A property literally named object with a string type: correctly left untouched (the key is walked as a value, has no matching type, so nothing is added).

One thing worth noting for whoever merges: _walk mutates schema in place and returns the same object, same as the old version did, so behavior for existing callers is unchanged there. Dropping jsonpath_ng entirely is a nice side benefit given this was its only use.

Looks correct and well-scoped to me.

[github-actions]

📚 Documentation preview: https://dottxt-ai.github.io/outlines/pr-preview/pr-1909/

Preview updates automatically with each commit.

[RobinPicard]

Thanks!