Quiz collegato alla guida · Difficile Livello
NVIDIA NeMo Guardrails Quiz
Tests the five NeMo rail stages, Colang and configuration roles, tool-call enforcement, evaluation, and limits of configured safeguards.
Domanda 1 di 8
Quale fase del rail NeMo è destinata a validare l'input utente prima che venga chiamato un modello?
Anteprima di ogni domanda8 domande
Le risposte e le spiegazioni vengono visualizzate mentre rispondi al quiz.
- Quale fase del rail NeMo è destinata a validare l'input utente prima che venga chiamato un modello?
- Recupero
- Execution
- Ingresso
- Uscita
- Dove agiscono i rail di recupero in un'applicazione aumentata dal recupero?
- During model pretraining
- On documents or chunks returned by retrieval
- Only on the final user-facing response
- After a tool has executed
- Quale tipo di rail controlla le chiamate a strumenti o funzioni e i loro argomenti?
- Ingresso
- Recupero
- Dialog
- Execution
- Quale funzione svolge un dialog rail in NeMo Guardrails?
- Validate image file codecs only
- Replace all retrieval logic
- Train the base LLM
- Constrain guided conversational flow across turns
- Come si inseriscono YAML e Colang nel toolkit documentato?
- YAML configures settings and rails; Colang expresses conversational flows and guardrail logic
- Colang stores the model weights
- YAML is only a log format
- Both are names for an output classifier
- Why should tool authorization be enforced at the execution boundary?
- Retrieved chunks execute automatically
- A response filter cannot undo an external action already performed
- Dialog rails always block every unsafe tool
- A user prompt is a secure permission token
- What determines whether a NeMo rail uses an LLM, external service, or custom code?
- The user’s browser
- A universal default shared by all installations
- The chosen configuration and implementation
- The name of the rail alone
- What does adding NeMo Guardrails establish about system safety?
- A configurable set of controls that still requires scenario testing and other safeguards
- That human review is never useful
- That prompts can no longer contain hostile instructions
- A guarantee against every attack and factual error