#1982·outlines

applications: Application does not detect output_type changes between calls

Author: harsh4vardhanCreated Aug 6, 2026Updated Aug 11, 2026

Application.__call__ caches the generator and re-creates it only when model changes (line 101):

python

if model != self.model:
    self.model = model
    self.generator = Generator(model, self.output_type)

output_type is a public instance attribute. If it is mutated between calls with the same model, the stale generator — built for the old type — is silently reused. The new constraint is never applied.

Minimum example:

python

from outlines.applications import Application
from unittest.mock import MagicMock, patch

with patch("outlines.applications.Generator") as MockGenerator:
    calls = []
    MockGenerator.side_effect = lambda m, ot: calls.append(ot) or MagicMock()

    app = Application(MagicMock(), output_type=int)
    model = MagicMock()

    app(model, {})        # generator built for int
    app.output_type = str
    app(model, {})        # same model — generator NOT rebuilt

    print(calls)          # [<class 'int'>] — str was never used

Fix: extend the invalidation condition to also detect output_type changes:

python

if model != self.model or self.output_type != self._cached_output_type:

(or track the output type that was used when the current generator was built).

Source: dottxt-ai/outlines

View original on GitHub View discussion on GitHub