ResponseCodec

Turns a typed response into the String a SemanticCache stores, and back.

The cache keeps text — that is what an LLM returns — but the second-most-common thing to cache is a structured output: a parsed JSON object, an extracted record, a tool-call. Rather than pull a serialization library onto the core classpath, kmemo ships the seam and you bring the codec, exactly as it does for Embedder and CacheStore. A kotlinx.serialization one is two lines:

val codec = object : ResponseCodec<Weather> {
    override fun encode(value: Weather) = Json.encodeToString(value)
    override fun decode(text: String) = Json.decodeFromString<Weather>(text)
}
val weather: Weather = cache.getOrPut(prompt, codec) { llm.extractWeather(it) }

encode and decode must round-trip: decode(encode(v)) has to equal v for anything you cache, or a hit will not match what a miss computed. decode receives exactly what encode produced for some earlier call (never arbitrary text), so it does not need to be defensive against foreign input.