autointent.modules.scoring.LLMDescriptionScorer#

class autointent.modules.scoring.LLMDescriptionScorer(generator_config=None, temperature=1.0, max_concurrent=15, max_per_second=10, max_retries=3, multilabel=False)#

Bases: autointent.modules.scoring._description.base.BaseDescriptionScorer

LLM-based description scorer for zero-shot intent classification using structured output.

This scorer uses a Large Language Model (LLM) with structured output to perform zero-shot intent classification. The LLM is prompted to categorize intent descriptions into three probability levels for each utterance: - Most probable (probability 1.0): Intents that are most likely to match the utterance - Promising (probability 0.5): Intents that are plausible but less confident - Unlikely (probability 0.0): All other intents (implicit)

This approach leverages the reasoning capabilities of LLMs to understand complex relationships between utterances and intent descriptions, potentially achieving high accuracy for nuanced classification tasks. However, it requires API access to LLM services and can be slower and more expensive than encoder-based methods.

Parameters:

  • generator_config (dict[str, Any] | None) – Configuration for the Generator instance (LLM model settings)

  • temperature (pydantic.PositiveFloat) – Temperature parameter for scaling classifier logits (default: 1.0)

  • max_concurrent (pydantic.PositiveInt | None) – Maximum number of concurrent async calls to LLM (default: 15)

  • max_per_second (pydantic.PositiveInt) – Maximum number of API calls per second for rate limiting (default: 10)

  • max_retries (pydantic.PositiveInt) – Maximum number of retry attempts for failed API calls (default: 3)

  • multilabel (bool) – Flag indicating classification task type

Example:#

from autointent.modules.scoring import LLMDescriptionScorer

# Initialize LLM scorer with OpenAI GPT
scorer = LLMDescriptionScorer(
    temperature=1.0,
    max_concurrent=10,
    max_per_second=5,
    max_retries=2
)

# Zero-shot classification with intent descriptions
descriptions = [
    "User wants to book or reserve transportation like flights, trains, or hotels",
    "User wants to cancel an existing booking or reservation",
    "User asks about weather conditions or forecasts"
]

# Fit using descriptions only (zero-shot approach)
scorer.fit([], [], descriptions)

# Make predictions on new utterances
test_utterances = ["Reserve a hotel room", "Delete my booking"]
probabilities = scorer.predict(test_utterances)
name = 'description_llm'#
generator_config#
max_concurrent = 15#
max_per_second = 10#
max_retries = 3#
classmethod from_context(context, temperature=1.0, generator_config=None, max_concurrent=15, max_per_second=10, max_retries=3)#
Parameters:

  • context (autointent.Context)

  • temperature (pydantic.PositiveFloat)

  • generator_config (dict[str, Any] | None)

  • max_concurrent (pydantic.PositiveInt | None)

  • max_per_second (pydantic.PositiveInt)

  • max_retries (pydantic.PositiveInt)

Return type:

LLMDescriptionScorer

get_implicit_initialization_params()#
Return type:

dict[str, Any]

clear_cache()#

Clear cached data in memory used by the generator.

Return type:

None

dump(path)#
Parameters:

path (str)

Return type:

None

classmethod load(path, embedder_config=None, cross_encoder_config=None)#
Parameters:

  • path (str)

  • embedder_config (autointent.configs._transformers.EmbedderConfig | None)

  • cross_encoder_config (autointent.configs._transformers.CrossEncoderConfig | None)

Return type:

LLMDescriptionScorer