LatencyEstimator

class LatencyEstimator

import { LatencyEstimator } from '@driftengine/ai';

In depth

The continuation watermark issues the next request when the current intent's remaining extent falls to this number. Measuring it rather than configuring it is what removes the tuning constant that would be wrong on every device, and wrong again the moment a consumer switched providers.

p90 rather than p50 or p99. At p50 half the continuations land late and the agent visibly drops to its floor mid-behaviour. At p99 the lead is long enough that the context goes stale for the tail. What would make this wrong: a provider whose latency is bimodal — a cache hit at 40ms and a miss at 3s — has a p90 that describes neither, and the answer there is a per-request estimate from the request's own shape, not a different percentile.

Constructor

constructor(capacity?: number, minSamples?: number)

Accessors

Methods

record(ms: number): void