Prevent Illegal Look-Around for OneOf in JSONSchema

[lapp0]

Fixes #823

This comment details the issues error: #823 (comment)

The reproduction code provided results in a json schema with OneOf[pets]:

class Model(BaseModel):
    pet: Union[Cat, Dog] = Field(..., discriminator='pet_type')

Before this PR: OneOf uses negative lookaheads to assert that only one schema member is included. This is illegal in interegular, more details available here: #456

After OneOf uses or-joined non-capturing groups which don't have the same issues with interegular.

[rlouf]

Looks good, thank you for the fix!

[ekagra-ranjan]

@lapp0 what is the difference bw oneOf and anyOf in this script? Both concat the subregex with |

[ekagra-ranjan]

Do you also think additional conditions like just checking for ^ and just checking for $ needs to be added to this conditon?

outlines/outlines/fsm/json_schema.py

Lines 248 to 249 in 7863f8e

	if pattern[0] == "^" and pattern[-1] == "$":
	return rf'(^"{pattern[1:-1]}"$)'

[ekagra-ranjan]

And how about adding array along with object to the types to ignore here?

outlines/outlines/fsm/json_schema.py

Lines 344 to 347 in 7863f8e

	regexes = [
	to_regex(resolver, {"type": t}, whitespace_pattern)
	for t in instance_type
	if t != "object"

Array like object requires many more parameters like items which is not provided in the list of types which is the reason why object is ignored I believe since properties cannot be mentioned when type is given as a list

[lapp0]

@ekagra-ranjan actually reviewing the test cases, I think anyOf is wrong. I'd appreciate a second set of eyes on this. Would you agree that

oneOf is correct (uses exclusive or), guarantees only one match

>>> print(re.fullmatch(r"((?:foo)|(?:bar))", "foo"))
<re.Match object; span=(0, 3), match='foo'>
>>> print(re.fullmatch(r"((?:foo)|(?:bar))", "foobar"))
None

anyOf is incorrect, allows only one match (I think this regression was introduced in https://github.com/outlines-dev/outlines/pull/552/files)

>>> print(re.fullmatch(r"((foo)|(bar))", "foo"))
<re.Match object; span=(0, 3), match='foo'>
>>> print(re.fullmatch(r"((foo)|(bar))", "foobar"))
None

allOf is somewhat correct, requires generation of all schemas, but enforces an order which it shouldn't:

    elif "allOf" in instance:
        subregexes = [
            to_regex(resolver, t, whitespace_pattern) for t in instance["allOf"]
        ]
        subregexes_str = [f"{subregex}" for subregex in subregexes]
        return rf"({''.join(subregexes_str)})"

[lapp0]

For your other two comments, I appreciate your insights into json schema -> regex. It's definitely a work in progress and it's great that you're looking deep into it's behavior and identifying problems. I think it's out of scope with this PR. Could you create a separate issue so others can provide input as well?