Session 2. Call a model through the adapter — Tue 15 Sep

Quick quiz (ungraded)

It inherits from LLMClient
Nothing inherits from it. LLMClient is a Protocol; inheritance is not how it is satisfied.
It has a complete(system, user) -> str method, and nothing else is required
Correct. Structural typing: the method is the contract, so a four-line fake is as valid a client as a vendor adapter.
It is registered in get_client
get_client maps configuration to a client. A class it has never heard of still satisfies the protocol.

The wording of the answer, the latency, and whether two runs agree
Correct. Those are the model's. Permissions, grading, redaction, retry budgets and failure behaviour are yours and must not move.
Which tools the model may call, because a better model can be trusted with more
A permission that depends on the model is not a permission. The lane must not widen it.
What happens on a failure, because a real provider fails differently
You will see different failures. What a failure does is your code's decision, and it stays the same.

The request goes out and the API returns 401
Nothing leaves the machine. get_client raises before the SDK is even constructed.
get_client raises ConfigError naming .env.example, and never prints the key
Correct. It fails closed at configuration time, and a message that echoed the key would put a secret in a notebook output.
It falls back to FakeLLM silently
The notebook preflight falls back and says so on its lane line. get_client itself raises; a silent downgrade would hide the missing key.

Nothing; let TimeoutError propagate so the caller sees the truth
The caller then handles two shapes and usually handles neither. ch02-e4 fails an escaped TimeoutError.
An AgentResult whose answer is flagged for review, cites nothing, and says the model did not respond in time
Correct. One shape either way, no citations for a call that never happened, and the flag downstream code routes on.
A best-effort answer from the fake, so the pipeline keeps moving
That is the one dishonest ending: a plausible answer produced by something other than the model you asked.

A FakeLLM, and the printed line is the only honest statement of that
Correct. The probe could not reach the server or the model, so the notebook keeps running on the fake and says so.
An OllamaClient that will raise on the first call
The lane check probes before it promises, so you get a client that works rather than one that fails later.
Undefined; the notebook stops there
The preflight never raises. A notebook that stops on a missing optional lane would block the required one.