AI Agents & GenAI Apps test (Advanced)

Sample questions

5 sample questions from the AI Agents & GenAI Apps test, with the answers. Samples are taken out of the live test, so you will not meet these when you sit it.

1. What is the primary behavior of MCP context compression in managing the context window during model processing?

   1. MCP context compression increases the context window size by adding placeholder tokens to accommodate future data.
   2. MCP context compression uses checksums to validate the integrity of the context window before processing.
   3. MCP context compression applies encryption to the context window to protect data during transmission.
   4. MCP context compression reduces the size of the context window by removing redundant or irrelevant information, ensuring that only essential data is retained for model processing.

   Answer: D. MCP context compression reduces the size of the context window by removing redundant or irrelevant information, ensuring that only essential data is retained for model processing.

   MCP context compression works by identifying and eliminating unnecessary or redundant information within the context, thereby optimizing the model's processing efficiency and reducing computational overhead. This allows the model to focus on the most relevant parts of the context, enhancing performance without sacrificing critical details.

2. How does a guardrail system typically resolve a conflict between two policies with differing restrictions?

   1. The system halts execution and waits for user input to resolve the conflict.
   2. The system randomly selects a policy to apply and ignores the conflict.
   3. The system applies the most restrictive policy and logs the conflict for resolution.
   4. The system applies the least restrictive policy and discards the conflict.

   Answer: C. The system applies the most restrictive policy and logs the conflict for resolution.

   In guardrail policy conflict resolution, the system prioritizes the most restrictive policy to ensure compliance and security. Conflicts are logged for manual review and resolution, ensuring that no less restrictive policy overrides a stricter one, which could compromise safety or regulatory requirements.

3. Which of the following accurately describes the behavior of LoRA adapters during model inference?

   1. LoRA adapters are only applied to the model's output layer.
   2. LoRA adapters are applied to the base model's weights during inference.
   3. LoRA adapters are used to replace the base model's weights entirely.
   4. LoRA adapters are applied during the initial model training phase.

   Answer: B. LoRA adapters are applied to the base model's weights during inference.

   LoRA (Low-Rank Adaptation) adapters are applied during inference to modify the base model's behavior without retraining the entire model. This allows for efficient fine-tuning with minimal additional parameters. The other options describe incorrect behaviors or misconceptions about how LoRA adapters operate.

4. Which statement accurately describes the application of LoRA training in a realistic project scenario?

   1. LoRA training is a method of fully retraining a pre-trained model from scratch using a smaller dataset for a specific task.
   2. LoRA training is a technique that replaces the model's weights entirely with new parameters derived from a smaller dataset, making it unsuitable for large models.
   3. LoRA training involves using high-rank matrices to increase the model's capacity for a new task, which increases training time and computational costs.
   4. LoRA training allows for efficient adaptation of large pre-trained models by introducing low-rank matrices to the model's weight layers, which significantly reduces the number of parameters that need to be fine-tuned compared to full model fine-tuning.

   Answer: D. LoRA training allows for efficient adaptation of large pre-trained models by introducing low-rank matrices to the model's weight layers, which significantly reduces the number of parameters that need to be fine-tuned compared to full model fine-tuning.

   LoRA training is a technique that enables the adaptation of large pre-trained models with minimal computational cost by adding low-rank matrices to the model's weights. This approach avoids the need to retrain the entire model, which is both computationally expensive and time-consuming. In a realistic project scenario, this is particularly useful when working with large language models where full model fine-tuning is impractical due to resource constraints. The other options either describe incorrect behavior or are not specific to LoRA training.

5. What is the behavior of a tool call with streaming responses when the model generates output?

   1. The response is streamed in chunks as the model generates output, allowing for real-time processing and feedback.
   2. The model waits for user input after each chunk of the response is sent.
   3. The tool call is executed only after the entire response is generated by the model.
   4. The response is sent in a single, unprocessed block after the model completes its generation.

   Answer: A. The response is streamed in chunks as the model generates output, allowing for real-time processing and feedback.

   When using tool calling with streaming responses, the model generates output incrementally, sending partial results as they are produced. This allows for real-time interaction and processing, unlike waiting for the full response before taking action. The other options describe behaviors that are not associated with streaming responses.