outlines

JsonSchema.__eq__ compares only the parsed schema JSON and omits whitespace_pattern, so schemas with different patterns are treated as identical and cache/deduplication can reuse the wrong logits processor. Fix equality to compare the pattern as well:

def __eq__(self, other):
    if not isinstance(other, JsonSchema):
        return False
    try:
        self_dict = json.loads(self.schema)
        other_dict = json.loads(other.schema)
        return (
            self_dict == other_dict
            and self.whitespace_pattern == other.whitespace_pattern
        )
    except json.JSONDecodeError:
        return (
            self.schema == other.schema
            and self.whitespace_pattern == other.whitespace_pattern
        )

Until patched, avoid caching processors by JsonSchema alone; use a cache key such as (schema.schema, schema.whitespace_pattern) or clear the relevant cache when the pattern changes.

Update scripts/gen_ref_pages.py so mkdocs-gen-files scans the new package location. Replace the old path root / "outlines" with root / "src" / "outlines". After the change, rebuild the docs with MkDocs and verify that pages such as site/api_reference/models/index.html are generated. The issue occurs because the src-layout migration moved the package from outlines/ to src/outlines/, leaving the old glob empty.

In src/outlines/types/json_schema_utils.py, json_schema_dict_to_dataclass assigns None as the default for non-required fields but does not wrap their type in Optional. Update the converter so non-required properties use Optional[typ]:

from typing import Optional

for property, details in properties.items():
    typ = schema_type_to_python(details, "dataclass")

    if property not in required:
        typ = Optional[typ]
        defaults[property] = None

    annotations[property] = typ

This makes an optional JSON Schema integer generate age: Optional[int] = None, matching the Pydantic converter and allowing null in structured generation. As a temporary workaround, manually declare non-required dataclass fields as Optional[T] = None, or use convert_to(schema, "pydantic") until the fix is applied.

The numeric value is converted to a regex without escaping, so the . in 1.5 acts as a regex wildcard. In src/outlines/types/dsl.py, change numeric term construction from Regex(str(value)) to Regex(re.escape(str(value))), ensuring re is imported:

import re

# Inside python_types_to_terms handling for int/float:
term = Regex(re.escape(str(value)))

This produces 1\.5, so Literal[1.5] matches only "1.5" and rejects "1x5". Until the fix is released, use an explicitly escaped regex such as Regex(r"1\.5") instead of a float literal constraint.

The ipv6 regex reuses the old IPv4 octet pattern, so suffixes such as ::ffff:01.1.1.1 or ::ffff:00.0.0.0 incorrectly match.

Fix it in src/outlines/types/__init__.py by replacing both occurrences of the old octet pattern in the ipv6 definition:

Old pattern:
25[0-5]|2[0-4][0-9]|1?[0-9]{1,2}

New pattern:
25[0-5]|2[0-4][0-9]|1[0-9][0-9]|[1-9]?[0-9]

This must be updated in both IPv4-mapped branches:
- ::ffff:a.b.c.d
- x:x:x:x::a.b.c.d

The new pattern accepts canonical octets such as 0, 1, 42, and 192, but rejects leading-zero forms such as 00, 01, and 007.

As a temporary workaround before patching, validate generated addresses with Python's standard library:

import ipaddress

def is_valid_ip(value: str) -> bool:
    try:
        ipaddress.ip_address(value)
        return True
    except ValueError:
        return False

After applying the regex fix, ::ffff:192.168.1.1 and 1:2:3:4::5.6.7.8 still match, while ::ffff:01.1.1.1, ::ffff:00.0.0.0, and 1:2:3:4::5.6.007.8 are rejected.