tactus 0.33.0__py3-none-any.whl → 0.34.1__py3-none-any.whl

tactus/stdlib/classify/primitive.py

@@ -0,0 +1,287 @@
+"""
+ClassifyPrimitive - Smart classification with built-in retry and validation.
+
+This primitive wraps the classification infrastructure to provide:
+- Automatic retry with conversational feedback
+- Output validation against valid classes
+- Confidence extraction from response or logprobs
+- Structured result format (value, confidence, explanation)
+
+The primitive supports multiple classification methods:
+- "llm" (default): LLM-based classification with retry logic
+- "fuzzy": String similarity based classification (coming soon)
+"""
+
+import logging
+from typing import Any, Dict
+
+from ..core.base import BaseClassifier, ClassifierFactory
+from ..core.models import ClassifierResult
+from .llm import LLMClassifier
+from .fuzzy import FuzzyMatchClassifier
+
+__all__ = ["ClassifyPrimitive", "ClassifyHandle", "ClassifierResult"]
+
+logger = logging.getLogger(__name__)
+
+# Register classifier methods with the factory
+ClassifierFactory.register("llm", LLMClassifier)
+ClassifierFactory.register("fuzzy", FuzzyMatchClassifier)
+
+
+class ClassifyHandle:
+    """
+    A reusable classifier handle for Lua interop.
+
+    This is a thin wrapper around a BaseClassifier that handles
+    Lua table conversion.
+
+    Created by Classify { ... } and can be called multiple times.
+    """
+
+    def __init__(
+        self,
+        classifier: BaseClassifier,
+        lua_table_from: Any = None,
+    ):
+        """
+        Initialize ClassifyHandle.
+
+        Args:
+            classifier: The underlying BaseClassifier instance
+            lua_table_from: Function to convert Python dicts to Lua tables
+        """
+        self._classifier = classifier
+        self.lua_table_from = lua_table_from
+
+        # Expose classifier attributes
+        self.classes = classifier.classes
+        self.target_classes = classifier.target_classes
+
+        # For test access
+        self._agent = getattr(classifier, "_agent", None)
+
+    def __call__(self, input_value: Any) -> ClassifierResult:
+        """
+        Classify the input.
+
+        Args:
+            input_value: Input text or dict with 'text' field
+
+        Returns:
+            ClassifierResult with value, confidence, explanation
+        """
+        # Extract text from input
+        if isinstance(input_value, dict):
+            text = input_value.get("text") or input_value.get("input") or str(input_value)
+        else:
+            text = str(input_value)
+
+        return self._classifier.classify(text)
+
+    def reset(self):
+        """Reset the classifier state."""
+        self._classifier.reset()
+
+    @property
+    def total_calls(self) -> int:
+        """Get total number of calls made."""
+        return getattr(self._classifier, "total_calls", 0)
+
+    @property
+    def total_retries(self) -> int:
+        """Get total number of retries."""
+        return getattr(self._classifier, "total_retries", 0)
+
+    def __repr__(self) -> str:
+        return f"ClassifyHandle(classifier={self._classifier})"
+
+
+class ClassifyPrimitive:
+    """
+    Smart classification primitive with retry logic.
+
+    Follows the Agent pattern - can be configured once and called multiple times,
+    or used as a one-shot classifier.
+
+    Example usage in Lua:
+        -- One-shot classification
+        result = Classify {
+            classes = {"Yes", "No"},
+            prompt = "Did the agent greet the customer?",
+            input = transcript
+        }
+
+        -- Reusable classifier
+        classifier = Classify {
+            classes = {"positive", "negative", "neutral"},
+            prompt = "What is the sentiment?"
+        }
+        result1 = classifier(text1)
+        result2 = classifier(text2)
+
+        -- With target classes for metrics
+        classifier = Classify {
+            classes = {"Yes", "No", "NA"},
+            target_classes = {"Yes"},
+            prompt = "Did the agent comply?"
+        }
+    """
+
+    def __init__(
+        self,
+        agent_factory: Any,
+        lua_table_from: Any = None,
+        registry: Any = None,
+        mock_manager: Any = None,
+    ):
+        """
+        Initialize ClassifyPrimitive.
+
+        Args:
+            agent_factory: Factory function to create Agent instances
+            lua_table_from: Function to convert Python dicts to Lua tables
+            registry: Optional registry for accessing mocks
+            mock_manager: Optional mock manager for testing
+        """
+        self.agent_factory = agent_factory
+        self.lua_table_from = lua_table_from
+        self.registry = registry
+        self.mock_manager = mock_manager
+
+    def __call__(self, config: Dict[str, Any]) -> Any:
+        """
+        Create a classifier from configuration.
+
+        This is called when Lua does: Classify { ... }
+
+        Args:
+            config: Classification configuration
+                - classes: List of valid classification values (required)
+                - prompt: Classification prompt/instruction (required for LLM)
+                - input: Optional input for one-shot classification
+                - method: Classification method ("llm" or "fuzzy", default: "llm")
+                - target_classes: Classes for precision/recall metrics (optional)
+                - max_retries: Maximum retry attempts (default: 3)
+                - temperature: Model temperature (default: 0.3)
+                - model: Model to use (optional)
+                - confidence_mode: "heuristic" or "none" (default: "heuristic")
+
+        Returns:
+            ClassifyHandle if no input provided (reusable)
+            dict if input provided (one-shot result)
+        """
+        # Convert Lua table to Python dict
+        config = self._lua_to_python(config)
+
+        # Debug: log the config
+        import logging
+
+        logger = logging.getLogger(__name__)
+        logger.info(f"[Classify] Received config: {config}")
+
+        method = config.get("method", "llm")
+
+        # Validate required fields based on method
+        if method == "llm":
+            classes = config.get("classes")
+            if not classes:
+                raise ValueError("Classify requires 'classes' field")
+            prompt = config.get("prompt")
+            if not prompt:
+                raise ValueError("Classify requires 'prompt' field")
+        elif method == "fuzzy":
+            # Fuzzy mode can use either 'expected' (binary) or 'classes' (multi-class)
+            if not config.get("expected") and not config.get("classes"):
+                raise ValueError(
+                    "Classify with method='fuzzy' requires either 'expected' "
+                    "(for binary Yes/No) or 'classes' (for multi-class matching)"
+                )
+        else:
+            # Unknown method - let factory handle validation
+            pass
+
+        # Create the classifier using the factory
+        classifier = self._create_classifier(config)
+
+        # Wrap in handle for Lua interop
+        handle = ClassifyHandle(
+            classifier=classifier,
+            lua_table_from=self.lua_table_from,
+        )
+
+        # If input is provided, do one-shot classification
+        input_text = config.get("input")
+        if input_text is not None:
+            result = handle(input_text)
+            return self._to_lua_table(result.to_dict())
+
+        return handle
+
+    def _create_classifier(self, config: Dict[str, Any]) -> BaseClassifier:
+        """
+        Create a classifier based on configuration.
+
+        Args:
+            config: Configuration dict
+
+        Returns:
+            BaseClassifier instance
+        """
+        method = config.get("method", "llm")
+
+        if method == "llm":
+            return LLMClassifier(
+                classes=config["classes"],
+                prompt=config["prompt"],
+                agent_factory=self.agent_factory,
+                target_classes=config.get("target_classes"),
+                max_retries=config.get("max_retries", 3),
+                temperature=config.get("temperature", 0.3),
+                model=config.get("model"),
+                confidence_mode=config.get("confidence_mode", "heuristic"),
+                name=config.get("name"),
+            )
+        else:
+            # Use the factory for other methods
+            # Add agent_factory to config for methods that need it
+            factory_config = {**config, "agent_factory": self.agent_factory}
+            return ClassifierFactory.create(factory_config)
+
+    def _lua_to_python(self, value: Any) -> Any:
+        """Convert Lua table to Python dict recursively."""
+        if value is None:
+            return None
+
+        try:
+            from lupa import lua_type
+
+            if lua_type(value) == "table":
+                # Check if it's an array (1-indexed sequential keys)
+                result = {}
+                max_int_key = 0
+                has_string_keys = False
+
+                for k, v in value.items():
+                    if isinstance(k, int):
+                        max_int_key = max(max_int_key, k)
+                    else:
+                        has_string_keys = True
+                    result[k] = self._lua_to_python(v)
+
+                # If all keys are sequential integers 1..n, convert to list
+                if not has_string_keys and max_int_key == len(result):
+                    return [result[i] for i in range(1, max_int_key + 1)]
+
+                return result
+            return value
+        except ImportError:
+            return value
+
+    def _to_lua_table(self, value: Any) -> Any:
+        """Convert Python value to Lua table."""
+        if self.lua_table_from is None:
+            return value
+        if isinstance(value, dict):
+            return self.lua_table_from(value)
+        return value

tactus/stdlib/core/__init__.py

@@ -0,0 +1,57 @@
+"""
+Tactus Standard Library - Core Module
+
+Provides abstract base classes, common models, and shared utilities
+used across all stdlib primitives.
+
+Base Classes:
+- BaseClassifier: ABC for all classification strategies
+- BaseExtractor: ABC for all extraction strategies
+
+Models:
+- ClassifierResult: Result from any classifier
+- ExtractorResult: Result from any extractor
+- EvaluationResult: Metrics from evaluation
+
+Utilities:
+- RetryWithFeedback: Retry logic with conversational feedback
+- extract_confidence: Confidence extraction heuristics
+- validate_output: Output validation
+"""
+
+from .base import (
+    BaseClassifier,
+    BaseExtractor,
+    ClassifierFactory,
+    ExtractorFactory,
+)
+from .models import (
+    ClassifierResult,
+    ExtractorResult,
+    EvaluationResult,
+    ClassifierConfig,
+    ExtractorConfig,
+)
+from .retry import RetryWithFeedback
+from .confidence import extract_confidence
+from .validation import validate_output
+
+__all__ = [
+    # Base classes
+    "BaseClassifier",
+    "BaseExtractor",
+    # Factories
+    "ClassifierFactory",
+    "ExtractorFactory",
+    # Result models
+    "ClassifierResult",
+    "ExtractorResult",
+    "EvaluationResult",
+    # Config models
+    "ClassifierConfig",
+    "ExtractorConfig",
+    # Utilities
+    "RetryWithFeedback",
+    "extract_confidence",
+    "validate_output",
+]

tactus/stdlib/core/base.py

@@ -0,0 +1,320 @@
+"""
+Abstract base classes for stdlib primitives.
+
+These ABCs define the interface that all classifiers and extractors must implement,
+enabling polymorphism and consistent behavior across different implementations.
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List, Optional, Type
+
+from .models import ClassifierResult, ExtractorResult, EvaluationResult
+
+
+class BaseClassifier(ABC):
+    """
+    Abstract base class for all classification strategies.
+
+    Subclasses must implement the `classify` method. This enables:
+    - LLMClassifier: Uses LLM with retry logic
+    - FuzzyMatchClassifier: Uses string similarity
+    - Custom classifiers: User-defined implementations
+
+    All classifiers return ClassifierResult for consistent API.
+
+    Example:
+        class MyClassifier(BaseClassifier):
+            def classify(self, input_text: str) -> ClassifierResult:
+                # Custom classification logic
+                return ClassifierResult(value="Yes", confidence=0.9)
+    """
+
+    # Configuration (set by subclasses)
+    classes: List[str] = []
+    target_classes: List[str] = []
+    name: Optional[str] = None
+
+    @abstractmethod
+    def classify(self, input_text: str) -> ClassifierResult:
+        """
+        Classify the input text and return a result.
+
+        Args:
+            input_text: The text to classify
+
+        Returns:
+            ClassifierResult with value, confidence, explanation, etc.
+        """
+        ...
+
+    def __call__(self, input_value: Any) -> ClassifierResult:
+        """
+        Make classifiers callable.
+
+        Handles both string input and dict input (for Lua interop).
+
+        Args:
+            input_value: Either a string or dict with 'text'/'input' key
+
+        Returns:
+            ClassifierResult
+        """
+        if isinstance(input_value, dict):
+            text = input_value.get("text") or input_value.get("input") or str(input_value)
+        else:
+            text = str(input_value)
+
+        return self.classify(text)
+
+    def reset(self) -> None:
+        """
+        Reset any internal state (e.g., conversation history).
+
+        Override in subclasses that maintain state.
+        """
+        pass
+
+    @classmethod
+    def evaluate(
+        cls,
+        classifier: "BaseClassifier",
+        test_data: List[Dict[str, Any]],
+        label_key: str = "label",
+        input_key: str = "text",
+    ) -> EvaluationResult:
+        """
+        Evaluate a classifier on test data.
+
+        Args:
+            classifier: The classifier instance to evaluate
+            test_data: List of dicts with input text and expected labels
+            label_key: Key for expected label in test data
+            input_key: Key for input text in test data
+
+        Returns:
+            EvaluationResult with accuracy, precision, recall, F1
+        """
+        from collections import defaultdict
+
+        predictions = []
+        labels = []
+        total_retries = 0
+        confidences = []
+        errors = []
+
+        for item in test_data:
+            text = item.get(input_key, "")
+            expected = item.get(label_key)
+
+            try:
+                result = classifier.classify(text)
+                predictions.append(result.value)
+                labels.append(expected)
+                total_retries += result.retry_count
+                if result.confidence is not None:
+                    confidences.append(result.confidence)
+            except Exception as e:
+                errors.append(f"Error on item: {str(e)}")
+                predictions.append("ERROR")
+                labels.append(expected)
+
+        # Calculate accuracy
+        correct = sum(pred == label for pred, label in zip(predictions, labels))
+        accuracy = correct / len(test_data) if test_data else 0.0
+
+        # Calculate confusion matrix
+        confusion: Dict[str, Dict[str, int]] = defaultdict(lambda: defaultdict(int))
+        for pred, label in zip(predictions, labels):
+            confusion[label][pred] += 1
+
+        # Calculate precision/recall for target classes
+        precision = None
+        recall = None
+        f1 = None
+
+        if classifier.target_classes:
+            # True positives: predicted target AND was target
+            tp = sum(
+                1
+                for pred, label in zip(predictions, labels)
+                if pred in classifier.target_classes and label in classifier.target_classes
+            )
+            # False positives: predicted target BUT was NOT target
+            fp = sum(
+                1
+                for pred, label in zip(predictions, labels)
+                if pred in classifier.target_classes and label not in classifier.target_classes
+            )
+            # False negatives: did NOT predict target BUT was target
+            fn = sum(
+                1
+                for pred, label in zip(predictions, labels)
+                if pred not in classifier.target_classes and label in classifier.target_classes
+            )
+
+            precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
+            recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
+            f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
+
+        return EvaluationResult(
+            accuracy=accuracy,
+            precision=precision,
+            recall=recall,
+            f1=f1,
+            confusion_matrix=dict(confusion),
+            total_samples=len(test_data),
+            total_retries=total_retries,
+            mean_confidence=sum(confidences) / len(confidences) if confidences else None,
+            errors=errors,
+        )
+
+
+class BaseExtractor(ABC):
+    """
+    Abstract base class for all extraction strategies.
+
+    Subclasses must implement the `extract` method. This enables:
+    - LLMExtractor: Uses LLM with retry logic
+    - SchemaExtractor: Uses structured schema parsing
+    - Custom extractors: User-defined implementations
+
+    All extractors return ExtractorResult for consistent API.
+
+    Example:
+        class MyExtractor(BaseExtractor):
+            def extract(self, input_text: str) -> ExtractorResult:
+                # Custom extraction logic
+                return ExtractorResult(fields={"name": "John", "age": 30})
+    """
+
+    # Configuration (set by subclasses)
+    fields: Dict[str, str] = {}
+    name: Optional[str] = None
+
+    @abstractmethod
+    def extract(self, input_text: str) -> ExtractorResult:
+        """
+        Extract structured data from the input text.
+
+        Args:
+            input_text: The text to extract from
+
+        Returns:
+            ExtractorResult with fields dict and validation info
+        """
+        ...
+
+    def __call__(self, input_value: Any) -> ExtractorResult:
+        """
+        Make extractors callable.
+
+        Handles both string input and dict input (for Lua interop).
+
+        Args:
+            input_value: Either a string or dict with 'text'/'input' key
+
+        Returns:
+            ExtractorResult
+        """
+        if isinstance(input_value, dict):
+            text = input_value.get("text") or input_value.get("input") or str(input_value)
+        else:
+            text = str(input_value)
+
+        return self.extract(text)
+
+    def reset(self) -> None:
+        """
+        Reset any internal state.
+
+        Override in subclasses that maintain state.
+        """
+        pass
+
+
+class ClassifierFactory:
+    """
+    Factory for creating classifiers based on configuration.
+
+    Supports registration of custom classifier types.
+    """
+
+    _registry: Dict[str, Type[BaseClassifier]] = {}
+
+    @classmethod
+    def register(cls, method: str, classifier_class: Type[BaseClassifier]) -> None:
+        """Register a classifier type."""
+        cls._registry[method] = classifier_class
+
+    @classmethod
+    def create(cls, config: Dict[str, Any], **kwargs) -> BaseClassifier:
+        """
+        Create a classifier from configuration.
+
+        Args:
+            config: Configuration dict with 'method' key
+            **kwargs: Additional kwargs passed to classifier constructor
+
+        Returns:
+            BaseClassifier instance
+
+        Raises:
+            ValueError: If method is not registered
+        """
+        method = config.get("method", "llm")
+
+        if method not in cls._registry:
+            available = ", ".join(cls._registry.keys())
+            raise ValueError(f"Unknown classifier method: '{method}'. Available: {available}")
+
+        classifier_class = cls._registry[method]
+        return classifier_class(config=config, **kwargs)
+
+    @classmethod
+    def available_methods(cls) -> List[str]:
+        """Get list of available classifier methods."""
+        return list(cls._registry.keys())
+
+
+class ExtractorFactory:
+    """
+    Factory for creating extractors based on configuration.
+
+    Supports registration of custom extractor types.
+    """
+
+    _registry: Dict[str, Type[BaseExtractor]] = {}
+
+    @classmethod
+    def register(cls, method: str, extractor_class: Type[BaseExtractor]) -> None:
+        """Register an extractor type."""
+        cls._registry[method] = extractor_class
+
+    @classmethod
+    def create(cls, config: Dict[str, Any], **kwargs) -> BaseExtractor:
+        """
+        Create an extractor from configuration.
+
+        Args:
+            config: Configuration dict with 'method' key
+            **kwargs: Additional kwargs passed to extractor constructor
+
+        Returns:
+            BaseExtractor instance
+
+        Raises:
+            ValueError: If method is not registered
+        """
+        method = config.get("method", "llm")
+
+        if method not in cls._registry:
+            available = ", ".join(cls._registry.keys())
+            raise ValueError(f"Unknown extractor method: '{method}'. Available: {available}")
+
+        extractor_class = cls._registry[method]
+        return extractor_class(config=config, **kwargs)
+
+    @classmethod
+    def available_methods(cls) -> List[str]:
+        """Get list of available extractor methods."""
+        return list(cls._registry.keys())