IReasoningModelDiagnosticSink

Interfaces · AiDotNet.Interfaces

Receives the record of every adjustment a reasoning-model profile makes to an outgoing request.

Known implementations: CollectingReasoningModelDiagnosticSink

For Beginners

When the library has to change your request so a newer “reasoning” model will accept it, it tells you. This interface is where those notices go. Write a class that implements it and, say, prints each notice to your log; hand it to the chat client and you will never be surprised by a setting that quietly stopped applying.

How It Works

A reasoning model rejects some of the sampling settings an ordinary chat model accepts, so the library removes them, renames one, and sometimes adds a deliberation level. Dropping a caller’s setting without saying so would make a run impossible to explain, so every edit is offered to this sink. Implement it to forward the record to whatever logging, metrics, or audit trail the host already uses; CollectingReasoningModelDiagnosticSink is a ready-made in-memory implementation for tests and short runs.

An implementation must be safe to call from several threads at once, because one chat client is normally shared by concurrent evolution workers. It must also not throw: the connectors treat a sink failure as a reporting problem and continue, but a sink that throws on every call turns a diagnostic path into a hot exception path. Keep the implementation cheap — it runs once per adjusted setting per request.

Methods

Method Summary
void Report(ReasoningModelDiagnostic diagnostic) Records one adjustment made to an outgoing request.