AI Architecture & the Feedback Loop

Question 1

The layer added immediately after the bare model API in most AI applications is:

Accepted Answer

Context construction — The first failure every project hits is that the model does not know your documents or today's facts, which is a context problem.

Question 2

Guardrails are enforced outside the model because:

Accepted Answer

A model cannot reliably separate instructions from the data it processes — That inability is exactly what prompt injection exploits, so model-side defences are probabilistic mitigations rather than boundaries.

Question 3

The real prompt-injection attack surface in a RAG application is usually:

Accepted Answer

Content pulled in by the retriever — A web page, PDF or email your retriever fetched is untrusted text arriving in the same channel as your instructions.

Question 4

Which is an output guardrail rather than an input guardrail?

Accepted Answer

Schema validation of the response — Schema validation inspects what the model produced. The other three all act on the request before it reaches the model.

Question 5

Capping input length and attachment count protects primarily against:

Accepted Answer

Unbounded cost and denial-of-wallet — Unbounded input is unbounded spend, and it is the easiest cost-exhaustion vector an AI endpoint exposes.

Question 6

Schema validation should fail closed rather than coerce because:

Accepted Answer

A hijacked or malformed response usually fails validation first — The parser is an accidental security control. Coercing a malformed response repairs exactly the signal you needed.

Question 7

The only hard boundary in an agent system is:

Accepted Answer

Least-privilege tool access — Every other layer is probabilistic. A capability the model does not hold cannot be misused however the prompt is manipulated.

Question 8

A gateway layer exists primarily to centralise:

Accepted Answer

Keys, rate limits, retries and fallbacks — One place holding those concerns is what makes a provider outage or deprecation a config change rather than an incident across the codebase.

Question 9

Explicit feedback such as thumbs up/down typically has:

Accepted Answer

Very low volume and a negative skew — People rate when annoyed. Low single-digit rates skewed negative are normal, which makes it a failure queue rather than a metric.

Question 10

Which implicit signal most directly provides a labelled correction?

Accepted Answer

The user edited the output before using it — The diff hands you the input, the model's attempt and the right answer — the shape of both an eval case and a training example.

Close the feedback loop

Test your knowledge

The layer added immediately after the bare model API in most AI applications is:

The short version

In more depth

AI Architecture & the Feedback Loop

The layers

Feedback signals

Designing the ask

Limits

Interview Questions

References & Further Reading

Related topics

Love this visualization?