AIMentor — AI Expertise Learning Platform

13 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
What is the main advantage of native function calling over the text-based ReAct protocol?
It uses fewer tokens in every caseThe model returns a structured, schema-conformant tool call instead of free text you must parseIt removes the need for a loopIt works without an LLM
Question 2easy
How is a tool described to the model for function calling?
As raw Python source codeAs a name, a description, and a JSON Schema for its parametersAs an embedding vectorAs a few-shot example only
Question 3medium
In the structured tool-call loop, what must each tool-result message reference?
The system promptThe originating call's tool_call_id, so the model pairs each result to its requestThe embedding of the tool outputThe user's IP address
Question 4medium
How does the loop know the agent is finished in native function calling?
The response contains no tool_calls (it returns plain content instead)A 'Final Answer:' string appearsThe temperature reaches 0The tool returns null
Question 5hard
A JSON schema guarantees arguments have the right types. Why still validate them?
Schemas are usually ignored by the modelSchema conformance is shape, not validity — the model can produce plausible-but-wrong values (negative price, invented id)Validation makes the call fasterBecause required fields are optional
Question 6medium
In the examples, what role does pydantic play?
It generates embeddingsIt performs runtime validation/coercion of tool arguments (types plus custom business rules)It calls the LLMIt stores vectors
Question 7medium
When can a model request multiple tool calls in a single turn?
Never — one call per turn is a hard limitWhen the calls are independent (e.g. weather in Paris AND Tokyo); you execute each and return one result per idOnly when using embeddingsOnly for the calculator tool
Question 8hard
A tool raises an exception (timeout, 404, bad value). What is the correct handling in an agent loop?
Let the exception propagate and crash the loopCatch it at the tool boundary and return the error as the observation so the model can recoverSilently return an empty stringImmediately stop the agent with a generic message
Question 9hard
Why should tool error messages be informative rather than a bare exception like 'KeyError'?
For prettier logs onlyThe model reads the error to decide its recovery, so 'no flights under €200' enables a far better next step than 'KeyError'The API rejects short error stringsIt reduces token cost
Question 10hard
Which are good practices for agent tool use? (Select all that apply)
Validate argument business rules before executingKey each tool result to its tool_call_idCatch tool exceptions and return them as observationsPass raw model-generated arguments straight into SQL or shell
Select all that apply.
Question 11medium
Within a tool's JSON schema, what does an `enum` on a property express?
The property is requiredThe set of allowed values for that field (e.g. unit: 'c' or 'f')The maximum string lengthThat the field is an embedding
Question 12medium
What data format (two words) is used to describe a tool's parameters to the model?
Question 13hard
In one word, what should you do with a tool exception so the agent can keep going: crash, or ____ it and return the error as an observation?

0 / 13 answered

Tool Use & Function Calling — Final Exam