confidence-router · packages

A threshold-based decision engine that takes a classification result with confidence scores and returns a `RoutingDecision` indicating whether to **route** (high confidence), **clarify** (ambiguous), or **fallback** (low confidence).

A pluggable classifier system for confidence-router, providing keyword matching, embedding similarity, and LLM-based classification as classes conforming to the `Classifier` interface, with a `ClassifierRegistry` for named registration and automatic fallback chains.

Core type definitions, error classes, configuration utilities, and the `DecisionEngine` for the confidence-router ecosystem. Exports TypeScript types (`Prediction`, `RoutingDecision`, `RouterConfig`, `Classifier` interface), a `RouterError` class with enumerated error types, a `DecisionEngine` class that evaluates confidence scores against route/clarify/fallback thresholds, and configuration helpers (`DEFAULT_CONFIG`, `validateConfig`, `mergeConfig`). Zero runtime dependencies.

A grid search optimizer that tunes `routeThreshold` and `fallbackThreshold` on any `RouterInterface` object to maximize F1 score against a labeled dataset, returning `OptimizedThresholds` with the best thresholds and full `EvaluationMetrics`.

Provides locale-aware clarification prompt generation for confidence-router, exposing `LanguageManager` and `PromptGenerator` classes that handle 47 built-in languages with RTL support, list separators, and conjunctions, falling back to English for unknown codes.