crestapps-core-ai-usage-metering

CrestApps.Core AI Usage Metering

Every client IAIClientFactory returns meters its own requests by wrapping the provider's client and reading the provider-agnostic UsageDetails on each response. That covers every provider, including custom IAIClientProvider implementations, with no provider code.

Guidelines

• Turn metering on with GeneralAIOptions.EnableAIUsageTracking = true. It is off by default.
• A record is written per request sent to the provider, not per user turn. Planning, query rewrites, vision calls, each tool-calling round trip, each embedding batch and each realtime response are separate records, so the request count equals billed requests.
• Metering sits below every middleware you add, so UseFunctionInvocation() yields one record per round trip and a cached response yields none.
• Label your own requests with UseUsageLabels(...) on the client pipeline, or an AIUsageScope around a call into a service that creates the client for you. Labels never reach the provider.
• A client you build without the factory is not metered until you wrap it with AIUsageMetering.Meter(...).

Enable Metering and Storage

using CrestApps.Core.AI.Models;

builder.Services.Configure<GeneralAIOptions>(options =>
{
    options.EnableAIUsageTracking = true;
});

Each record goes to every registered IAICompletionUsageObserver. The default IAICompletionUsageService is one of them and saves through IAICompletionUsageStore. The YesSql chat-session stores (AddCoreAIChatSessionStoresYesSql(), which calls AddCoreAICompletionUsageStoresYesSql()) and the EntityCore chat-session stores (AddCoreAIChatSessionStoresEntityCore()) register a store. A record whose scope was already disposed when the response finished is dropped rather than failing the answer.

What a Record Holds