| """Output schemas are application-owned, never inferred from user prompt text."""
|
| from functools import lru_cache
|
|
|
|
|
| def obj(properties, required=None):
|
| return {'type':'object', 'properties':properties,
|
| 'required':list(properties) if required is None else required,
|
| 'additionalProperties':False}
|
|
|
|
|
| def array(items, minimum=0):
|
| return {'type':'array','items':items,'minItems':minimum,'maxItems':6}
|
|
|
|
|
| INDEX = {'type':'integer','enum':[0,1,2,3,4,5]}
|
| BOOL = {'type':'boolean'}
|
| SCHEMAS = {
|
| 'claims':obj({
|
| 'status':{'type':'string','enum':['answer','insufficient_evidence','needs_clarification']},
|
| 'claims':array(obj({'text':{'type':'string','minLength':1,'maxLength':450},
|
| 'source_id':{'type':'integer'},'excerpt_id':{'type':'integer'}}))}),
|
| 'support':obj({'verdicts':array(obj({'claim_index':INDEX, 'supported':BOOL,
|
| 'reason_code':{'type':'string','enum':['supported','extra_facts','reversed_meaning',
|
| 'wrong_region','not_answering_question','wrong_language','uncertain']}},
|
| required=['claim_index','supported']), minimum=1)}),
|
| 'translation':obj({'steps':array(obj({'index':INDEX,
|
| 'text':{'type':'string','minLength':1,'maxLength':900}}), minimum=1)}),
|
| 'translation_review':obj({'verdicts':array(obj({'index':INDEX,
|
| 'correct_language':BOOL,'same_meaning':BOOL}), minimum=1)}),
|
| }
|
|
|
|
|
| class StructuredPrompt(str):
|
| def __new__(cls, prompt, kind):
|
| if kind not in SCHEMAS:
|
| raise ValueError('Unknown output schema')
|
| instance = super().__new__(cls, prompt)
|
| instance.kind = kind
|
| return instance
|
|
|
| def __getnewargs__(self):
|
|
|
| return str(self), self.kind
|
|
|
|
|
| @lru_cache(maxsize=2)
|
| def _tokenizer_data(tokenizer):
|
| from lmformatenforcer.integrations.transformers import build_token_enforcer_tokenizer_data
|
| return build_token_enforcer_tokenizer_data(tokenizer)
|
|
|
|
|
| def prefix_constraint(tokenizer, prompt):
|
| """Fail explicitly if formatting support is missing; do not silently run unconstrained."""
|
| from lmformatenforcer import JsonSchemaParser
|
| from lmformatenforcer.integrations.transformers import build_transformers_prefix_allowed_tokens_fn
|
| if not isinstance(prompt, StructuredPrompt):
|
| raise ValueError('Generation requires an application-owned output schema.')
|
| return build_transformers_prefix_allowed_tokens_fn(_tokenizer_data(tokenizer), JsonSchemaParser(SCHEMAS[prompt.kind]))
|
|
|
