Type Alias AdaptiveTokenEstimatorOptions

Type Declaration

• OptionalmaxCorrection?: number

  Upper clamp for the applied correction factor. Must be

  = minCorrection. Default: 2.

• OptionalmaxModels?: number

  Maximum distinct models tracked. When a new model would exceed this, the oldest-inserted entry is evicted. Bounds memory in the face of unbounded / attacker-controlled model strings. Must be a positive integer. Default: 64.

• OptionalminCorrection?: number

  Lower clamp for the applied correction factor. Guards admission against a run of anomalous observations collapsing estimates toward zero. Must be > 0. Default: 0.5.

• OptionalminSamples?: number

  Observations required for a model before its correction is applied to estimates. Until then the estimator returns the uncorrected base estimate. Must be a positive integer. Default: 5.

• Optionaloverrides?: Record<string, number>

  Exact-model ratio overrides, forwarded to the underlying model-aware estimator (same semantics as the two-argument form of createModelAwareTokenEstimator).

• Optionalsmoothing?: number

  EWMA smoothing factor in (0, 1]. Each observation moves the per-model correction factor by smoothing * (observed - current). Higher adapts faster but is noisier. Default: 0.2.