dotnet-microsoft-extensions-ai

Define tools with AIFunction and AIFunctionFactory, then wrap your client with FunctionInvokingChatClient to handle the invocation loop automatically. Validate tool arguments explicitly and keep tool registration scoped to the current conversation to control token costs.

Applications and services should reference Microsoft.Extensions.AI, which includes middleware, caching, telemetry, and function invocation. Reference only the Abstractions package when authoring provider libraries or reusable connectors at the contract layer.

For stateless providers, resend the full message history each turn. For stateful providers, propagate ConversationId from ChatResponse into ChatOptions instead of replaying history. Check the response ConversationId to detect which mode applies.

Yes, Ollama-backed clients implement the same IChatClient abstraction, enabling low-cost local prototyping and offline development. You still own chat history replay, latency, and model-quality tradeoffs when using local models.

Escalate to Agent Framework when you need agent threads, multi-agent orchestration, durable execution, workflows, or remote agent hosting. Microsoft.Extensions.AI covers provider abstraction and middleware composition, not autonomous orchestration.

Use the Microsoft.Extensions.AI.Evaluation packages: Quality evaluators for relevance and groundedness, NLP evaluators like BLEU and F1 for cheap baseline checks, and Safety evaluators for harmful content. Cache responses and publish reports so CI runs stay comparable.