Oracle Cloud Infrastructure Generative AI Professional 1Z0-1127-25 1Z0-1127-25 Questions 301–375 | Page 5/11

301

MCQhard

An AI application uses chain-of-thought prompting to solve math problems. The team observes that the model sometimes arrives at correct answers but with flawed reasoning. Which technique can specifically help by aggregating multiple reasoning paths?

A.Role prompting

B.Self-consistency prompting

C.Tree-of-thought prompting

D.Few-shot prompting

AnswerB

Self-consistency samples multiple chain-of-thought outputs and chooses the most consistent answer, effectively filtering flawed reasoning.

Why this answer

Self-consistency generates several reasoning chains and selects the most common answer, improving robustness even if some chains have flawed steps.

Full explanation →

302

MCQmedium

A company is deploying a large language model for a customer service chatbot. The model needs to understand industry-specific jargon and maintain low latency. Which approach best balances these requirements?

A.Employ retrieval-augmented generation (RAG) with a general model

B.Rely solely on prompt engineering with a general model

C.Use a large general-purpose LLM with zero-shot prompting

D.Fine-tune a small open-source LLM on domain-specific data

AnswerD

Fine-tuning adapts the model to jargon and a smaller model keeps latency low.

Why this answer

Fine-tuning a small open-source LLM on domain-specific data is the best approach because it adapts the model to understand industry-specific jargon while keeping the model small enough to maintain low latency. Unlike larger models, a fine-tuned small model can run efficiently on local hardware, reducing inference time and avoiding the overhead of external API calls or large model sizes.

Exam trap

Oracle often tests the misconception that larger models always perform better or that RAG alone solves domain adaptation, ignoring the latency and efficiency trade-offs that make fine-tuning a smaller model the optimal choice for production systems with strict response time requirements.

How to eliminate wrong answers

Option A is wrong because retrieval-augmented generation (RAG) with a general model still relies on a general model that may not inherently understand industry-specific jargon, and the retrieval step adds latency, which conflicts with the low-latency requirement. Option B is wrong because relying solely on prompt engineering with a general model does not embed domain-specific knowledge into the model weights, so the model may still misinterpret or fail to generate accurate responses for niche jargon, and it often requires longer prompts that increase latency. Option C is wrong because a large general-purpose LLM with zero-shot prompting has high inference latency due to its size and lacks domain-specific training, making it unsuitable for both understanding jargon and meeting low-latency constraints.

Full explanation →

303

MCQmedium

A developer is using OCI Generative AI Service to generate product descriptions. The outputs are often too generic and lack brand-specific tone. The developer has a small set of 20 high-quality example descriptions. What is the most efficient approach to improve output quality?

A.Fine-tune a base model on the 20 examples.

B.Use few-shot prompting by including the 20 examples in the prompt.

C.Use a more detailed system prompt describing the brand tone.

D.Use chain-of-thought prompting to guide the model step by step.

AnswerB

Few-shot prompting leverages examples without retraining, ideal for small datasets.

Why this answer

Few-shot prompting is the most efficient approach when you have a small set of high-quality examples (20 in this case). It allows the model to infer the desired tone and style directly from the provided examples without requiring any training or fine-tuning, which would be inefficient and potentially ineffective with such a small dataset. In OCI Generative AI Service, few-shot prompting leverages the model's in-context learning capability to adapt its output to the brand-specific tone.

Exam trap

Oracle often tests the misconception that fine-tuning is always the best approach for customization, but candidates overlook the fact that with very small datasets (like 20 examples), few-shot prompting is more practical and efficient than fine-tuning.

How to eliminate wrong answers

Option A is wrong because fine-tuning a base model on only 20 examples is inefficient and unlikely to produce reliable results; fine-tuning typically requires hundreds to thousands of high-quality examples to avoid overfitting and to meaningfully adjust model weights. Option C is wrong because a more detailed system prompt describing the brand tone, while helpful, is less effective than providing concrete examples; the model may still produce generic outputs without specific stylistic references. Option D is wrong because chain-of-thought prompting is designed to improve reasoning and step-by-step logic, not to adapt tone or style; it does not address the core issue of generating brand-specific product descriptions.

Full explanation →

304

MCQeasy

A developer is using a large language model to generate code snippets. The model often produces code that is syntactically correct but functionally incorrect. What is the most effective way to improve the functional correctness of the generated code?

A.Provide few-shot examples of correct code in the prompt.

B.Increase the temperature parameter to generate more creative solutions.

C.Ask the model to only output syntactically valid code.

D.Set max_tokens to a very high value to allow the model more room to think.

AnswerA

Few-shot examples help the model understand the expected output.

Why this answer

Providing few-shot examples of correct code in the prompt directly demonstrates the desired functional behavior to the model. This technique, known as few-shot prompting, grounds the model's output in concrete examples, significantly improving the likelihood that the generated code will be functionally correct by aligning the model's pattern completion with the intended logic, not just syntax.

Exam trap

Oracle often tests the misconception that increasing model parameters like temperature or max_tokens can improve output quality, when in fact these parameters control randomness and length, not functional correctness, which is best addressed through prompt engineering techniques like few-shot learning.

How to eliminate wrong answers

Option B is wrong because increasing the temperature parameter makes the model's output more random and creative, which typically reduces functional correctness by increasing the chance of generating plausible but incorrect logic. Option C is wrong because asking the model to only output syntactically valid code does not address functional correctness; the model already generates syntactically valid code by default, and this instruction does not guide it toward correct logic. Option D is wrong because setting max_tokens to a very high value does not improve reasoning quality; it only allows longer outputs, which can actually increase the risk of generating more irrelevant or incorrect code without improving functional correctness.

Full explanation →

305

MCQmedium

A team is building an agent using LangChain that needs to perform calculations and search the web for current information. Which combination of tools and agent type should they use?

A.Zero-shot agent with only a calculator tool

B.Conversational agent with a search tool and memory

C.ReAct agent with a search tool only

D.ReAct agent with a calculator tool and a search tool

AnswerD

ReAct agents combine reasoning and acting; a calculator handles arithmetic, and a search tool retrieves live web data.

Why this answer

A ReAct agent reasons step-by-step and uses tools like a calculator for math and a search tool for web queries. The ReAct framework is designed for such tool-use tasks. The other options either omit necessary tools or use an incorrect agent type.

Full explanation →

306

MCQeasy

What is the primary difference between pre-training and fine-tuning in the context of large language models?

A.Pre-training trains from scratch, fine-tuning updates all weights on a new dataset

B.Pre-training uses a smaller dataset, fine-tuning uses a larger dataset

C.Pre-training produces embeddings, fine-tuning produces text generation

D.Pre-training is unsupervised, fine-tuning is always supervised

AnswerA

Pre-training involves training from random initialization on a large corpus; fine-tuning starts from pre-trained weights and updates them on a smaller dataset.

Why this answer

Pre-training trains a model on a large, general corpus to learn language representations; fine-tuning adapts the pre-trained model to a specific task or domain using a smaller labeled dataset.

Full explanation →

307

MCQhard

A team fine-tunes an embedding model for a legal document RAG system but observes low retrieval recall. Which technique is most likely to improve recall?

A.Use a smaller batch size

B.Use hard negative mining during training

C.Reduce the learning rate

D.Increase the number of fine-tuning epochs

AnswerB

Hard negatives force the model to differentiate between similar but irrelevant documents, improving retrieval discrimination.

Why this answer

Hard negative mining exposes the model to challenging negatives during training, which sharpens the embedding space and improves recall.

Full explanation →

308

MCQmedium

A developer is using OCI Generative AI with a Cohere Command model for text generation. They want the output to be more creative and diverse, but still relevant. Which sampling strategy should they use?

A.Temperature sampling (temperature > 1)

B.Top-k sampling

C.Top-p (nucleus) sampling

D.Greedy decoding

AnswerC

Top-p sampling dynamically chooses the set of tokens with cumulative probability p, balancing creativity and relevance by adapting to the model's confidence.

Why this answer

Top-p (nucleus) sampling selects from the smallest set of tokens whose cumulative probability exceeds p. It adapts to the model's confidence, allowing diversity while maintaining relevance. Greedy decoding is deterministic, temperature scales all probabilities, top-k fixes the number of candidates, and beam search explores multiple sequences but tends to produce safe outputs.

Full explanation →

309

MCQhard

An organization requires low-latency inference for a custom fine-tuned model deployed on OCI Generative AI. The model must be isolated from other tenants. Which infrastructure choice meets these requirements?

A.Use the shared infrastructure endpoint for the fine-tuned model

B.Deploy the model on OCI Data Science model deployment

C.Provision a dedicated AI cluster with the required model units

D.Use an on-premises GPU server and call OCI GenAI via API gateway

AnswerC

A dedicated AI cluster provides isolated compute resources, ensuring low latency and no interference from other workloads.

Why this answer

A dedicated AI cluster provides isolated, low-latency inference for custom models. Shared infrastructure does not guarantee isolation and may have higher latency due to contention.

Full explanation →

310

MCQmedium

Refer to the exhibit. A developer runs the command and immediately tries to use the endpoint. The application fails with an error indicating the endpoint is not active. What is the most likely reason?

A.The model ID is not available in us-ashburn-1

B.The purpose parameter is misspelled

C.The endpoint is in provisioning state and not yet ready

D.The compartment ID is incorrect

AnswerC

Endpoints take time to provision; using them immediately fails.

Why this answer

When a developer creates a new model deployment endpoint in OCI, the endpoint enters a 'provisioning' state and can take several minutes to become active. Attempting to use the endpoint immediately after creation will result in an error indicating the endpoint is not ready, as the underlying infrastructure and model loading must complete before inference requests are accepted.

Exam trap

Oracle often tests the asynchronous nature of OCI resource creation, where candidates mistakenly assume the endpoint is immediately usable after the create command returns a success response.

How to eliminate wrong answers

Option A is wrong because the model ID being unavailable in us-ashburn-1 would cause a different error during deployment creation, not an 'endpoint not active' error after the command runs. Option B is wrong because a misspelled 'purpose' parameter would cause a validation error at command submission time, not a runtime error when trying to use the endpoint. Option D is wrong because an incorrect compartment ID would result in an authorization or not-found error during the create command, not a post-creation 'endpoint not active' error.

Full explanation →

311

MCQmedium

A team uses Cohere's `rerank` endpoint after initial retrieval to improve result quality. What is the main benefit of reranking?

A.It generates new embeddings for chunks

B.It combines multiple queries

C.It reorders chunks by relevance to the query

D.It reduces the number of retrieved chunks

AnswerC

Reranking improves the ordering so the most relevant appear first.

Why this answer

The rerank endpoint takes the initial set of retrieved chunks and re-scores them based on their semantic relevance to the query, producing a reordered list where the most contextually appropriate chunks appear first. This improves the quality of the input fed to the LLM, leading to more accurate and coherent generated responses without altering the embeddings or the retrieval count.

Exam trap

Oracle often tests the distinction between retrieval and reranking, and the trap here is that candidates confuse reranking with reducing the number of chunks (Option D) or assume it modifies embeddings (Option A), when in fact it only reorders based on deeper relevance scoring.

How to eliminate wrong answers

Option A is wrong because reranking does not generate new embeddings; it uses the existing embeddings from the initial retrieval step to compute relevance scores. Option B is wrong because reranking operates on a single query against a set of retrieved chunks, not by combining multiple queries. Option D is wrong because reranking does not reduce the number of retrieved chunks; it reorders the existing set, and the number of chunks passed to the LLM remains the same unless explicitly truncated elsewhere.

Full explanation →

312

MCQeasy

Which of the following is an example of a prompt template?

A."Write a product description for {product_name} targeting {audience}."

B."You are a helpful assistant."

C."Translate the following English text to French: 'Hello'"

D."Explain the concept of gravity."

AnswerA

Placeholders make it a template.

Why this answer

A template uses placeholders like {product_name} that are filled at runtime, making prompts reusable and dynamic.

Full explanation →

313

MCQmedium

A customer support company uses Cohere Command on OCI to answer user queries. They have enabled grounding with a knowledge base of product manuals. However, for about 20% of queries, the model provides incorrect product recommendations that are not in the manuals. The team has verified the knowledge base is up to date. What is the most likely cause and solution?

A.The model's temperature is too high, causing creative responses. Lower temperature to 0.

B.The model is hallucinating; switch to a larger model.

C.The query phrasing may not match the knowledge base; improve the retrieval system or use query rewriting.

D.The grounding settings are too restrictive; increase the number of retrieved documents.

AnswerC

Correct: Query mismatch causes retrieval of irrelevant content, leading to incorrect recommendations.

Why this answer

The most likely cause is that the query phrasing does not align well with the knowledge base content, leading to retrieval of irrelevant or no documents. Even with an up-to-date knowledge base, if the query is phrased differently from the manual's text, grounding may fail to retrieve the correct information. Improving the retrieval system (e.g., using semantic search) or implementing query rewriting can bridge this gap, enabling the model to use the appropriate context and provide accurate recommendations.

Option A might help if temperature were high, but the core issue is retrieval, not creativity. Option B (switching to a larger model) does not address the retrieval problem. Option D (increasing retrieved documents) could introduce more noise if the retrieved documents are already irrelevant.

Full explanation →

314

MCQeasy

Which component of the Transformer architecture allows the model to focus on different parts of the input sequence when generating each output token?

A.Self-attention mechanism

B.Positional encoding

C.Feed-forward network

D.Layer normalization

AnswerA

Self-attention allows each token to attend to all other tokens.

Why this answer

Self-attention computes attention scores between all pairs of positions, enabling the model to weigh the importance of different input tokens.

Full explanation →

315

MCQmedium

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

A.Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store

B.Train a custom model from scratch on the policy documents each month

C.Use a larger foundation model with a longer context window and paste all documents into each prompt

D.Fine-tune a base LLM on the policy documents monthly

AnswerA

RAG retrieves relevant document chunks at query time, ensuring the chatbot always answers from the latest uploaded documents without any model retraining.

Why this answer

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

Full explanation →

316

MCQeasy

A developer wants the model to generate a bullet list of steps to troubleshoot a network issue. Which output format specification should be included in the prompt?

A.Use a few-shot example with numbered steps

B.Provide a table format

C.Set temperature to 0

D.Request 'bullet list' in the output format

AnswerD

Specifying 'bullet list' in the output format prompts the model to generate each step as a bullet point.

Why this answer

Explicitly requesting a bullet list in the output format specification is the most direct way to achieve the desired structure.

Full explanation →

317

MCQeasy

A user wants to access the OCI Generative AI service programmatically. Which credential method is recommended for use in a production application running on OCI Compute?

A.API signing keys

B.Instance principal

C.User password and OCID

D.Resource principal

AnswerB

Instance principal dynamically obtains credentials via instance metadata service.

Why this answer

Instance principal authentication is the recommended method for production applications running on OCI Compute because it allows the application to authenticate with OCI services without managing or embedding any credentials. The OCI Compute instance assumes a dynamic group and IAM policy that grants it permissions, and the SDK automatically handles token exchange via the instance metadata service, eliminating the need for long-lived secrets.

Exam trap

Oracle often tests the distinction between instance principal (for Compute instances) and resource principal (for serverless or managed services), leading candidates to confuse the two or to incorrectly select API signing keys as the 'most secure' option without considering operational overhead.

How to eliminate wrong answers

Option A is wrong because API signing keys are long-lived secrets that must be securely stored and rotated, which adds operational overhead and risk in a production environment. Option C is wrong because user passwords and OCIDs are intended for interactive console login, not programmatic API access, and they cannot be used with the OCI SDK or CLI for service calls. Option D is wrong because resource principal is used for serverless functions like OCI Functions or for resources like OCI Object Storage buckets, not for a Compute instance running a custom application.

Full explanation →

318

MCQeasy

What is the primary goal of chain-of-thought prompting?

A.Prevent the model from hallucinating

B.Make the output more creative

C.Reduce output token count

D.Encourage step-by-step reasoning to improve accuracy on multi-step problems

AnswerD

Correct: it breaks down reasoning into intermediate steps.

Why this answer

Chain-of-thought encourages the model to perform step-by-step reasoning before arriving at an answer, improving performance on complex tasks.

Full explanation →

319

MCQmedium

A.Fine-tune a base LLM on the policy documents monthly

B.Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store

C.Use a larger foundation model with a longer context window and paste all documents into each prompt

D.Train a custom model from scratch on the policy documents each month

AnswerB

RAG retrieves relevant document chunks at query time, ensuring the chatbot always answers from the latest uploaded documents without any model retraining.

Why this answer

Full explanation →

320

Multi-Selecthard

A developer is using tree-of-thought prompting to solve a complex planning problem. Which THREE characteristics are typical of tree-of-thought?

Select 3 answers

A.It explores multiple reasoning branches in parallel

B.It generates a single linear chain of reasoning

C.It allows backtracking to explore alternative branches

D.It always returns a single final answer via majority vote

E.It uses evaluation or heuristics to prune unpromising branches

AnswersA, C, E

Correct: branches represent different reasoning paths.

Why this answer

Tree-of-thought explores multiple reasoning branches in parallel, uses evaluation or heuristics to prune unpromising branches, and can backtrack to explore alternatives. It does not follow a single linear path or guarantee a single answer without voting.

Full explanation →

321

MCQeasy

What is the primary purpose of setting the 'top-k' parameter in an LLM?

A.To truncate the output after k tokens

B.To control the creativity of the output by scaling logits

C.To penalize tokens that have already appeared in the generated text

D.To limit the next token selection to the k most probable tokens

AnswerD

Top-k sampling selects only from the top k tokens by probability, ignoring the long tail of low-probability tokens.

Why this answer

Top-k restricts the model to sample only from the k most likely next tokens, reducing the chance of unusual or off-topic words. Top-p is nucleus sampling; temperature controls randomness; frequency penalty reduces repetition.

Full explanation →

322

MCQhard

Refer to the exhibit. A developer has set this policy to allow an OCI Data Science session to generate embeddings. However, the API call returns a 403 Forbidden. Which of the following is likely missing?

A.The policy needs a 'where request.region != ...' condition

B.The policy should include 'in tenancy' instead of compartment

C.The service requires 'manage' permission instead of 'use'

D.The dynamic group does not include the Data Science session

AnswerD

The session must be matched by a rule in the dynamic group for the policy to apply.

Why this answer

The 403 Forbidden error indicates that the Data Science session does not have the necessary permissions to call the API. In OCI, policies grant permissions to dynamic groups, not directly to resources. The dynamic group must include the Data Science session (e.g., by matching the session's OCID or resource type) for the policy to apply.

Without this, the session is not recognized as a principal, and the API call is denied.

Exam trap

Oracle often tests the distinction between policy syntax correctness and dynamic group membership, leading candidates to focus on permission levels or scope when the real issue is that the resource (Data Science session) is not a member of the dynamic group referenced in the policy.

How to eliminate wrong answers

Option A is wrong because adding a 'where request.region != ...' condition would restrict the policy to specific regions, but the core issue is that the session is not authorized at all, not that it's in the wrong region. Option B is wrong because 'in tenancy' would broaden the scope to all compartments, but the policy already targets a specific compartment; the problem is the dynamic group membership, not the scope. Option C is wrong because 'use' permission is sufficient for generating embeddings (which is a read/use operation), and 'manage' would be excessive; the error is not about permission level but about the principal not being authorized.

Full explanation →

323

Multi-Selectmedium

Which two actions are required when deploying a custom fine-tuned model using the OCI Generative AI service? (Choose two.)

Select 2 answers

A.Configure an API Gateway for the model endpoint

B.Register the model in the OCI Data Science Model Catalog

C.Set up a load balancer for the deployment

D.Upload model artifacts to OCI Object Storage

E.Create a dedicated AI cluster

AnswersD, E

Model artifacts must be stored in Object Storage before deployment.

Why this answer

Deploying a custom fine-tuned model in OCI Generative AI requires the model artifacts (e.g., weights, configuration files) to be stored in OCI Object Storage. The service pulls the artifacts from a designated bucket during deployment. Option E is correct because a dedicated AI cluster must be created to host the model, providing the necessary compute resources for inference.

Exam trap

The trap here is that candidates confuse the OCI Data Science Model Catalog (used for ML model lifecycle management) with the Generative AI service's own model registration, leading them to incorrectly select Option B.

Full explanation →

324

Multi-Selecthard

Which THREE parameters can be adjusted to reduce repetition in generated text? (Choose three.)

Select 3 answers

A.presence_penalty

B.top_k

C.max_tokens

D.frequency_penalty

E.temperature

AnswersA, B, D

Penalizes tokens that have appeared at all.

Why this answer

A is correct because presence_penalty directly penalizes tokens that have already appeared in the generated text, reducing the likelihood of repeating the same concepts or phrases. This parameter works by subtracting a fixed penalty from the logits of tokens that are already present in the sequence, making the model less likely to reuse them.

Exam trap

Oracle OCI often tests the distinction between parameters that control output length (max_tokens) versus those that control repetition (presence_penalty, frequency_penalty), and candidates mistakenly think temperature reduces repetition when it actually controls randomness.

Full explanation →

325

MCQeasy

Refer to the exhibit. What is the primary reason the response is incomplete?

A.The temperature is not set.

B.The model-id is incorrect.

C.The max-tokens limit is too low.

D.The prompt is too short.

AnswerC

Setting max-tokens to 100 restricts the output length, causing truncation.

Why this answer

The response is incomplete because the max-tokens limit is too low, causing the model to truncate its output before completing the full answer. When the token budget is exhausted, the generation stops mid-sentence or mid-thought, leaving the response unfinished regardless of prompt length or other parameters.

Exam trap

Oracle often tests the distinction between parameters that affect output quality (temperature, top_p) versus those that constrain output length (max_tokens, stop sequences), and the trap here is that candidates mistake a short prompt or missing temperature for the cause of truncation when the real culprit is the token budget.

How to eliminate wrong answers

Option A is wrong because the temperature parameter controls randomness in token selection, not the length or completeness of the response; a missing temperature would default to 1.0 and still allow full output. Option B is wrong because the model-id identifies which LLM to use (e.g., gpt-3.5-turbo or cohere.command-text-v14) and does not affect whether the response is truncated; an incorrect model-id would either fail to load or produce different output, not an incomplete one. Option D is wrong because a short prompt can still yield a complete response; prompt length influences context and relevance, but the max-tokens parameter is the direct limiter of output length.

Full explanation →

326

Multi-Selecthard

An OCI practitioner is comparing BERTScore with traditional n-gram metrics (ROUGE, BLEU) for evaluating summarization. Which THREE statements about BERTScore are true?

Select 3 answers

A.BERTScore uses pre-trained contextual embeddings from BERT

B.BERTScore does not require a reference text

C.BERTScore computes similarity based on exact n-gram overlap

D.BERTScore typically has higher correlation with human judgment than ROUGE

E.BERTScore is more robust to paraphrasing than ROUGE or BLEU

AnswersA, D, E

It leverages BERT's embeddings to compute semantic similarity.

Why this answer

BERTScore uses contextual embeddings, captures semantic similarity, and correlates better with human judgment than n-gram metrics.

Full explanation →

327

MCQeasy

What is the primary purpose of setting a low temperature (e.g., 0.1) when generating text with an LLM?

A.Make responses more deterministic and focused

B.Increase creativity and diversity of responses

C.Enable the model to use external tools

D.Reduce the maximum number of tokens generated

AnswerA

Correct: low temperature concentrates probability on high-likelihood tokens.

Why this answer

Temperature controls randomness; low values make the model deterministic and focused on the most likely tokens, which is ideal for factual tasks.

Full explanation →

328

MCQmedium

An organization wants to deploy a model that can summarize long financial reports (5000+ tokens) without losing context. Which model architecture is best suited for this requirement?

A.Encoder-decoder model (e.g., T5)

B.Mixture-of-experts model

C.Decoder-only model (e.g., GPT)

D.Encoder-only model (e.g., BERT)

AnswerA

Encoder-decoder architecture excels at summarization and can handle long inputs via the encoder.

Why this answer

Encoder-decoder models like T5 or BART are designed for sequence-to-sequence tasks such as summarization, and can handle long inputs with their encoder.

Full explanation →

329

MCQeasy

What is the main advantage of using chain-of-thought (CoT) prompting over standard few-shot prompting for complex reasoning tasks?

A.CoT reduces the number of tokens needed in the prompt

B.CoT allows the model to use any tools available in the environment

C.CoT improves performance on multi-step reasoning tasks by generating step-by-step explanations

D.CoT eliminates the need for any examples in the prompt

AnswerC

CoT explicitly models the reasoning chain, helping the model reach correct conclusions.

Why this answer

CoT encourages the model to articulate intermediate reasoning steps, which improves accuracy on tasks that require multi-step logic. Few-shot provides examples but not the reasoning process.

Full explanation →

330

MCQhard

A company deploys a fine-tuned model on an OCI Generative AI dedicated AI cluster. After deployment, they observe high latency during peak hours. The cluster has only one replica. Which action would most effectively reduce latency without increasing cost unnecessarily?

A.Increase the number of replicas to 10.

B.Enable auto-scaling with a maximum of 3 replicas.

C.Switch to a larger base model.

D.Move to a serverless deployment model.

AnswerB

Auto-scaling adjusts to demand; a max of 3 provides headroom without waste.

Why this answer

Enabling auto-scaling with a maximum of 3 replicas (Option B) is the most effective action because it dynamically adds replicas during peak hours to handle increased load, reducing latency, while limiting the maximum to 3 prevents unnecessary cost overruns. This balances performance and cost, unlike a fixed large replica count or switching models, which either wastes resources or fails to address the root cause of insufficient compute capacity.

Exam trap

Oracle often tests the misconception that more replicas always reduce latency, but the trap here is that candidates may choose Option A (10 replicas) without considering cost efficiency, while the correct answer requires balancing performance with cost constraints via auto-scaling.

How to eliminate wrong answers

Option A is wrong because increasing replicas to 10 would significantly raise costs without proportional latency benefits, as the cluster likely doesn't need that many replicas during non-peak hours, leading to idle resource waste. Option C is wrong because switching to a larger base model would increase inference latency and cost due to higher computational requirements, exacerbating the problem rather than solving it. Option D is wrong because moving to a serverless deployment model on OCI Generative AI would introduce cold-start latency and unpredictable scaling behavior, and it may not support fine-tuned models or dedicated cluster features, potentially increasing latency and cost.

Full explanation →

331

MCQmedium

A company is deploying a large language model in a customer-facing chatbot. The model's responses must be both accurate and safe. Which combination of techniques should be employed?

A.Use only a system prompt instructing the model to be accurate and safe.

B.Use retrieval-augmented generation (RAG) for factual accuracy and a content safety filter for safe outputs.

C.Use a high temperature for creativity and a safety classifier for blocking toxic outputs.

D.Fine-tune the model on all historical chat logs and use a high temperature.

AnswerB

RAG improves accuracy; safety filter ensures safety.

Why this answer

RAG grounds the model's responses in a verified external knowledge base, reducing hallucinations and improving factual accuracy, while a content safety filter (e.g., a classifier or guardrail) actively blocks toxic or unsafe outputs before they reach the user. This combination addresses both accuracy and safety independently, unlike a single system prompt which is easily bypassed.

Exam trap

Oracle often tests the misconception that a single technique (like a system prompt or fine-tuning) can simultaneously guarantee both accuracy and safety, when in practice they require separate, complementary mechanisms.

How to eliminate wrong answers

Option A is wrong because a system prompt alone is a static instruction that can be overridden by user input or model behavior, providing no enforcement mechanism for accuracy or safety. Option C is wrong because a high temperature increases randomness and creativity, which is counterproductive for accuracy and can amplify unsafe outputs; a safety classifier is a partial solution but does not address factual grounding. Option D is wrong because fine-tuning on all historical chat logs may introduce biases, errors, or unsafe patterns from the data, and a high temperature further degrades reliability.

Full explanation →

332

Multi-Selectmedium

Which THREE are common prompt failures that can degrade model performance?

Select 3 answers

A.Using too few tokens in the output

B.Prompt injection vulnerabilities that allow malicious input to override instructions

C.Ambiguous instructions that can be interpreted in multiple ways

D.Insufficient context for the model to answer correctly

E.Setting temperature too high

AnswersB, C, D

Prompt injection can hijack the model's behavior.

Why this answer

Ambiguous instructions, insufficient context, and prompt injection are well-known failure modes.

Full explanation →

333

Multi-Selectmedium

Which TWO measures can help reduce the risk of generating toxic or unsafe content when using OCI Generative AI Service?

Select 2 answers

A.Use few-shot prompting with examples that demonstrate safe and appropriate responses.

B.Disable model monitoring and logging to reduce overhead.

C.Increase the temperature parameter to make output more deterministic.

D.Fine-tune the model on a large dataset without any safety filtering.

E.Enable the built-in content filtering features provided by OCI Generative AI Service.

AnswersA, E

Safe examples help steer the model toward desired behavior.

Why this answer

Few-shot prompting provides the model with explicit examples of safe, appropriate responses, which helps steer the model's behavior toward desired outputs and reduces the likelihood of generating toxic or unsafe content. This technique leverages in-context learning to align the model's responses with the provided examples, making it a practical measure for content safety.

Exam trap

Oracle often tests the misconception that increasing temperature or disabling monitoring improves safety, when in fact these actions increase randomness and reduce oversight, respectively.

Full explanation →

334

Multi-Selectmedium

A developer is building a LangChain application that uses OCI Generative AI service. They want to implement streaming responses from the LLM to improve user experience. Which TWO actions are necessary to enable streaming?

Select 2 answers

A.Use a StreamingStdOutCallbackHandler or custom callback to handle stream events

B.Increase the chunk_size to speed up streaming

C.Disable memory to avoid buffering the stream

D.Use async LangChain syntax for streaming

E.Set the 'streaming' parameter to True when invoking the ChatOCIGenAI model

AnswersA, E

Why this answer

LangChain's streaming architecture requires a callback handler (like `StreamingStdOutCallbackHandler` or a custom one) to process tokens as they are emitted from the LLM. Without a callback, the stream events are generated but not consumed, so the application cannot output partial results in real time.

Exam trap

The trap here is that candidates confuse the `streaming` parameter (which tells the model to use the streaming API) with the callback handler (which actually processes the stream), assuming one alone is sufficient, but both are required for a complete streaming implementation.

Full explanation →

335

MCQeasy

A prompt engineer wants to generate a JSON object from an LLM response. Which prompt component is MOST critical to ensure the output is valid JSON?

A.Context and background

B.Task instruction

C.Examples (few-shot)

D.Output format specification

AnswerD

Explicitly stating 'Output as JSON' with the desired structure is the most direct way to get valid JSON.

Why this answer

Explicitly specifying the output format (e.g., 'Provide the answer as a JSON object with keys ...') is essential for structured output. The other components are important for other aspects but not for format compliance.

Full explanation →

336

MCQeasy

What is the primary benefit of using a Dedicated AI Cluster over On-Demand serving for deploying generative AI models on OCI?

A.Higher throughput and lower latency due to reserved capacity

B.No need to manage model versions

C.Automatic scaling to zero when not in use

D.Lower cost for variable workloads

AnswerA

Reserved capacity minimizes resource contention, improving performance.

Why this answer

A Dedicated AI Cluster provides reserved compute capacity on OCI, ensuring consistent high throughput and low latency for generative AI inference workloads. Unlike On-Demand serving, which shares resources and can suffer from contention or cold starts, a Dedicated AI Cluster guarantees that GPU resources are always available for your model, eliminating variability in response times.

Exam trap

Oracle often tests the misconception that Dedicated AI Clusters are cheaper for variable workloads, when in fact their fixed-cost model makes them optimal for steady-state, high-volume inference, not spiky or unpredictable traffic.

How to eliminate wrong answers

Option B is wrong because managing model versions is a separate concern handled by model registries and deployment pipelines, not a benefit specific to Dedicated AI Clusters. Option C is wrong because Dedicated AI Clusters are always-on and do not scale to zero; automatic scaling to zero is a feature of serverless or On-Demand serving to reduce costs when idle. Option D is wrong because Dedicated AI Clusters incur fixed costs for reserved capacity, making them more expensive for variable workloads compared to On-Demand serving, which charges per-usage and can scale down.

Full explanation →

337

MCQeasy

Which of the following best describes the role of a Retriever in a LangChain RAG pipeline?

A.It splits documents into smaller chunks

B.It loads documents from various sources like PDF or HTML

C.It encodes documents into dense vector representations

D.It fetches relevant document chunks from a vector store based on a query

AnswerD

A Retriever's primary function is to retrieve relevant documents from a store using similarity search or other methods.

Why this answer

A Retriever is responsible for fetching relevant documents from a vector store based on a query. It abstracts the search logic so that downstream components (like LLMChain) can use the retrieved context. Document loaders handle input, embeddings convert text to vectors, and splitters divide documents.

Full explanation →

338

MCQhard

A developer is using the OCI Generative AI Chat API with a system prompt to guide the assistant's behavior. They notice that after a few turns, the assistant starts ignoring the system instructions. What is the MOST likely cause?

A.The temperature is set too high, causing the model to ignore instructions

B.The system prompt (preamble) is only sent once and may be truncated as the conversation grows

C.The model is not fine-tuned to follow system prompts

D.The conversation history is not being sent with subsequent requests, so the model loses context

AnswerB

In the Chat API, the preamble is sent only at the start; if the conversation history exceeds the context window, the preamble may be dropped, causing the model to lose instructions.

Why this answer

The preamble is sent only on the first turn; it may be truncated or lost as conversation history grows. The developer should inject the preamble in each request or manage history carefully.

Full explanation →

339

MCQmedium

A.Train a custom model from scratch on the policy documents each month

B.Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store

C.Use a larger foundation model with a longer context window and paste all documents into each prompt

D.Fine-tune a base LLM on the policy documents monthly

AnswerB

RAG retrieves relevant document chunks at query time, ensuring the chatbot always answers from the latest uploaded documents without any model retraining.

Why this answer

Full explanation →

340

MCQmedium

A team is fine-tuning an LLM on OCI Generative AI for a domain-specific task. They have a dataset of 10,000 labeled examples. What is a best practice to avoid catastrophic forgetting during fine-tuning?

A.Increase the learning rate to speed up adaptation.

B.Use only the new domain-specific data for fine-tuning.

C.Reduce the number of training epochs to the minimum.

D.Include a small percentage of general-domain data in the training mix.

AnswerD

General data acts as a regularizer to maintain base knowledge.

Why this answer

Catastrophic forgetting occurs when a fine-tuned model loses previously learned general knowledge. By including a small percentage (e.g., 5–10%) of general-domain data in the training mix, the model retains its broad capabilities while adapting to the new domain-specific task. This technique, often called 'replay' or 'experience replay,' is a standard practice in continual learning for LLMs.

Exam trap

Oracle often tests the misconception that fine-tuning should exclusively use the new dataset, whereas the best practice is to blend in general data to preserve prior knowledge.

How to eliminate wrong answers

Option A is wrong because increasing the learning rate can cause the model to overfit to the new domain data and accelerate forgetting, not prevent it. Option B is wrong because using only domain-specific data removes all exposure to general knowledge, which is the primary cause of catastrophic forgetting. Option C is wrong because reducing epochs to the minimum may prevent the model from learning the new task adequately, but it does not address the retention of general knowledge; the model can still forget if the new data dominates the gradient updates.

Full explanation →

341

Multi-Selecteasy

Which TWO are advantages of using LoRA for fine-tuning?

Select 2 answers

A.Requires less GPU memory

B.Guarantees higher accuracy

C.Reduces number of trainable parameters

D.Increases model size

E.Improves inference speed

AnswersA, C

Fewer trainable parameters means lower memory usage during training.

Why this answer

LoRA (Low-Rank Adaptation) reduces GPU memory requirements because it freezes the original model weights and injects trainable low-rank matrices into specific layers. This means only a tiny fraction of parameters need gradients and optimizer states, drastically lowering memory consumption during fine-tuning compared to full fine-tuning.

Exam trap

Oracle often tests the misconception that reducing trainable parameters automatically improves inference speed, but LoRA's memory and parameter savings apply only to training, not to inference latency.

Full explanation →

342

MCQhard

A data science team is using OCI Data Science to fine-tune a model. They notice that training jobs are failing due to out-of-memory errors on the notebook session. What should they do to resolve this?

A.Enable autoscaling on the notebook session.

B.Use OCI Data Flow instead.

C.Switch to a larger notebook session shape.

D.Reduce the batch size in the training script.

AnswerC

A larger shape provides more memory, resolving OOM issues.

Why this answer

Out-of-memory errors during training on a notebook session indicate that the current shape's memory capacity is insufficient for the model or data being processed. Switching to a larger notebook session shape directly increases available RAM and compute resources, resolving the memory constraint without altering the training logic or infrastructure type.

Exam trap

Oracle often tests the misconception that autoscaling or reducing batch size can fix memory issues in a single-node notebook session, but the correct approach is to match the compute shape to the workload's memory requirements.

How to eliminate wrong answers

Option A is wrong because autoscaling adjusts the number of compute instances horizontally, not the memory of a single notebook session; it does not prevent OOM errors caused by insufficient per-instance memory. Option B is wrong because OCI Data Flow is a serverless Spark-based service for big data processing, not designed for fine-tuning deep learning models, and migrating would require rewriting the training pipeline. Option D is wrong because reducing batch size can mitigate memory usage but does not address the root cause of an undersized notebook session shape; it may also degrade training convergence or performance, and the question asks for a resolution to the failing jobs, not a workaround.

Full explanation →

343

MCQmedium

An LLM-based application must comply with data privacy regulations by not memorizing personally identifiable information (PII). Which technique best reduces memorization of PII?

A.Use a larger model with more parameters

B.Decrease the temperature during inference

C.Train with differential privacy

D.Increase the number of training epochs

AnswerC

Differential privacy bounds the influence of any single data point, reducing memorization.

Why this answer

Differential privacy (DP) is the correct technique because it directly limits the model's ability to memorize training data, including PII, by adding calibrated noise to the gradient updates during training. This ensures that the model's parameters do not encode specific individual records, providing a formal mathematical guarantee against memorization. Other options like model size, temperature, or training epochs do not address the root cause of memorization in the training process.

Exam trap

Oracle often tests the misconception that inference-time parameters like temperature or model size affect training data memorization, when in fact memorization is a training-phase phenomenon that must be addressed during training itself.

How to eliminate wrong answers

Option A is wrong because increasing model parameters generally increases the model's capacity to memorize training data, making PII leakage more likely, not less. Option B is wrong because temperature controls the randomness of output token sampling during inference, not the memorization of training data; it has no effect on whether PII is stored in the model weights. Option D is wrong because increasing training epochs typically leads to overfitting and greater memorization of training examples, including PII, as the model sees the data more times.

Full explanation →

344

MCQhard

A financial services company has deployed a custom fine-tuned model using OCI Generative AI service on a dedicated AI cluster for automated report generation. They use a Python application that sends prompts via the OCI SDK. Recently, they started seeing 429 Too Many Requests errors intermittently. The dedicated cluster has 2 replicas and the application is making about 100 requests per second. The cluster's documented throughput is 50 requests per second per replica. The company has not set up any throttling limits. What is the most likely cause of the 429 errors?

A.The application is exceeding the cluster's replica capacity.

B.The OCI SDK version is outdated.

C.The API calls are not authenticated.

D.The model's context window is too small for the prompts.

AnswerA

The cluster is at maximum throughput; bursts push over the limit.

Why this answer

The dedicated AI cluster has 2 replicas, each supporting 50 requests per second, for a total capacity of 100 requests per second. The application is sending exactly 100 requests per second, which matches the cluster's maximum throughput. However, without any throttling limits configured, the service will still enforce its built-in rate limits at the replica level, and any transient burst or uneven distribution across replicas can trigger 429 Too Many Requests errors.

The most likely cause is that the application is operating at the cluster's replica capacity boundary, leading to intermittent throttling.

Exam trap

The trap here is that candidates may assume the total cluster capacity (100 req/s) is the limiting factor, but the 429 errors occur because each replica enforces its own 50 req/s limit, and uneven request distribution or bursts can cause individual replicas to be overwhelmed.

How to eliminate wrong answers

Option B is wrong because an outdated OCI SDK version would typically cause compatibility issues or missing features, not 429 errors, which are specifically about rate limiting. Option C is wrong because unauthenticated API calls would result in 401 Unauthorized or 403 Forbidden errors, not 429 Too Many Requests. Option D is wrong because a model's context window size affects the length of prompts or responses, not the rate at which requests are processed; 429 errors are unrelated to context window constraints.

Full explanation →

345

MCQhard

A financial services company is concerned about data privacy when using OCI Generative AI service for processing sensitive customer data. They want to ensure that their data is not used to improve the model and is encrypted at rest and in transit. Which combination of OCI features should they implement?

A.Use OCI Object Storage buckets with encryption to store prompts and responses

B.Provision a dedicated endpoint, configure data privacy opt-out, and use OCI Vault for encryption keys

C.Deploy the model in an OCI Data Science project with a private endpoint

D.Use the on-demand API with default encryption

AnswerB

Dedicated endpoints allow data isolation; Vault provides customer-managed keys; privacy opt-out prevents use for training.

Why this answer

It addresses all three requirements: a dedicated endpoint ensures network isolation and encryption in transit, data privacy opt-out prevents OCI from using customer data for model improvement, and OCI Vault integration allows customers to manage their own encryption keys for data at rest, meeting the financial services company's strict data privacy and encryption needs.

Exam trap

Oracle often tests the misconception that encryption alone (at rest or in transit) is sufficient for data privacy, when in fact the critical requirement for preventing model improvement is the explicit data privacy opt-out mechanism, which is a separate control from encryption.

How to eliminate wrong answers

Option A is wrong because OCI Object Storage encryption only protects data at rest, not data in transit, and it does not prevent OCI from using prompts and responses for model training or improvement. Option C is wrong because deploying a model in an OCI Data Science project with a private endpoint provides network isolation but does not include a data privacy opt-out mechanism to prevent data from being used for model improvement, nor does it inherently enforce customer-managed encryption keys for the Generative AI service. Option D is wrong because the on-demand API with default encryption uses OCI-managed keys, which does not give the customer control over encryption keys, and it does not include a data privacy opt-out to prevent data usage for model improvement.

Full explanation →

346

MCQmedium

A company is deploying a RAG system for internal document search using OCI OpenSearch as the vector store. Users report that queries about recent policy changes return no results, even though the new policies were ingested. Which configuration is most likely missing?

A.The query should use a hybrid search combining keyword and vector.

B.The embeddings must be normalized before indexing.

C.The vector search index must have a refresh interval set to immediate.

D.The ingestion pipeline should use a text-splitting chunker.

AnswerC

Without immediate refresh, new documents may not be visible in search results.

Why this answer

OCI OpenSearch by default has a refresh interval (e.g., 1 second), which means recently ingested documents may not appear in search results immediately. Setting the refresh interval to 'immediate' forces the index to refresh after every write, making new documents searchable right away. Option A (hybrid search) improves relevance but does not affect document availability.

Option B (normalizing embeddings) is important for cosine similarity but not for search availability. Option D (text-splitting chunker) is a preprocessing step for ingestion but does not control when documents become searchable.

Full explanation →

347

Multi-Selecteasy

A developer is building a RAG pipeline with LangChain. They have loaded PDF documents with PDFLoader. Which TWO steps must they perform before indexing the documents into a vector store?

Select 2 answers

A.Translate the documents into English

B.Split the documents into chunks using a text splitter

C.Encrypt the documents with AES-256

D.Embed the document chunks using an embedding model

E.Compress the documents with gzip

AnswersB, D

Splitting ensures each chunk fits the model's token limit and improves retrieval granularity.

Why this answer

Documents loaded from PDFs must be split into smaller chunks (using a text splitter) and then embedded into vector representations before they can be indexed in a vector store.

Full explanation →

348

MCQeasy

Which OCI Generative AI parameter controls the diversity of generated text by increasing the probability of less likely tokens?

A.frequency_penalty

B.top_k

C.temperature

D.top_p

AnswerC

Higher temperature increases randomness by scaling logits before softmax.

Why this answer

The temperature parameter in OCI Generative AI controls the randomness of token selection by scaling the logits before applying the softmax function. A higher temperature (e.g., > 1.0) increases the probability of less likely tokens, making the output more diverse and creative, while a lower temperature (e.g., < 1.0) sharpens the distribution toward the most likely tokens, producing more deterministic and conservative text.

Exam trap

The OCI Generative AI exam often tests the distinction between temperature and top_p/top_k by asking which parameter directly scales probabilities to favor less likely tokens, leading candidates to confuse sampling strategies (top_p/top_k) with the logit scaling mechanism (temperature).

How to eliminate wrong answers

Option A is wrong because frequency_penalty reduces the likelihood of tokens that have already appeared in the generated text, thereby discouraging repetition but not directly increasing the probability of less likely tokens. Option B is wrong because top_k limits the sampling pool to the k most likely next tokens, which can reduce diversity by excluding less likely tokens entirely. Option D is wrong because top_p (nucleus sampling) selects from the smallest set of tokens whose cumulative probability exceeds a threshold p, which also restricts diversity by cutting off the long tail of low-probability tokens.

Full explanation →

349

Multi-Selecthard

A developer is building a multi-turn chatbot using the OCI Generative AI Chat API. Which THREE parameters or features should they configure to maintain coherent conversation history?

Select 3 answers

A.Temperature

B.Stop sequences

C.System prompt

D.Multi-turn conversation history

E.Preamble override

AnswersC, D, E

Defines the assistant's persona and behavior.

Why this answer

System prompt sets the assistant's behavior. Preamble override can customize the initial context. Multi-turn conversation history is managed by sending previous messages in the API call.

Temperature and stop sequences affect generation but not history management.

Full explanation →

350

MCQmedium

A company wants to deploy a private instance of a large language model on OCI for sensitive data processing. What is the recommended approach?

A.Use OCI Data Science with a publicly accessible model.

B.Use the OCI Generative AI public endpoint with data encryption.

C.Use OCI Dedicated AI Clusters for a private endpoint.

D.Use third-party model hosting outside OCI.

AnswerC

Dedicated AI Clusters offer isolated compute with private networking, meeting security and compliance needs.

Why this answer

OCI Dedicated AI Clusters provide a fully isolated, private endpoint for deploying large language models, ensuring that sensitive data never traverses the public internet. This approach meets the requirement for private inference with no data leaving the customer's tenancy, unlike public endpoints or third-party hosting.

Exam trap

Oracle often tests the misconception that encryption alone (Option B) is sufficient for private deployment, but the trap is that encryption does not eliminate the need for network isolation when processing sensitive data in a shared infrastructure environment.

How to eliminate wrong answers

Option A is wrong because OCI Data Science with a publicly accessible model exposes the model endpoint to the internet, violating the requirement for private, sensitive data processing. Option B is wrong because the OCI Generative AI public endpoint, even with data encryption, still routes traffic through OCI's shared infrastructure and public IP space, which does not guarantee the isolation needed for sensitive data. Option D is wrong because third-party model hosting outside OCI would require data to leave the OCI tenancy, breaking the requirement for a private deployment within OCI.

Full explanation →

351

Multi-Selecthard

Which THREE are known challenges when deploying large language models in production?

Select 3 answers

A.Bias in training data perpetuating stereotypes

B.High computational cost for inference

C.Hallucination of plausible but incorrect information

D.Fast inference speed due to parallelization

E.Low memory footprint

AnswersA, B, C

Models can reflect and amplify biases from training data.

Why this answer

Large language models (LLMs) are trained on vast, unfiltered internet text corpora that inherently contain societal biases. These biases are learned and can be amplified during inference, leading to outputs that perpetuate harmful stereotypes, which is a well-documented production challenge.

Exam trap

Oracle often tests the distinction between known challenges (bias, cost, hallucination) and desirable properties (fast inference, low memory) that are actually false for LLMs, trapping candidates who confuse optimization goals with current limitations.

Full explanation →

352

MCQhard

A company is building a document summarization pipeline using OCI Generative AI. They need to summarize thousands of legal documents efficiently. Which approach minimizes cost while maintaining quality?

A.Use the OCI Generative AI Summarisation API with a prebuilt model

B.Fine-tune a model on legal documents and deploy on a dedicated cluster

C.Use the Chat API with a system prompt instructing summarization

D.Use the Embedding API to vectorize documents and then use a custom summarization algorithm

AnswerA

The Summarisation API is purpose-built for summarization, uses on-demand pricing, and is cost-efficient for large volumes.

Why this answer

The Summarisation API is designed for efficient, cost-effective summarization. Dedicated clusters are expensive for batch workloads, and embedding-based approaches add complexity and cost.

Full explanation →

353

MCQmedium

An organization needs to select a tokenisation algorithm for a multilingual LLM that will process English, Chinese, and Korean text efficiently. Which tokenisation method is BEST suited for this requirement?

A.WordPiece

B.Byte-Pair Encoding (BPE)

C.SentencePiece

D.Character-level tokenisation

AnswerC

SentencePiece treats the input as a raw byte stream and does not require pre-tokenisation, making it ideal for multilingual corpora including CJK languages.

Why this answer

SentencePiece is language-agnostic, works directly on raw text without requiring pre-tokenisation (e.g., whitespace splitting), and handles languages like Chinese and Korean where word boundaries are not obvious. BPE and WordPiece typically require pre-tokenisation, making them less suitable for CJK languages.

Full explanation →

354

Multi-Selectmedium

A data scientist is building a RAG pipeline on OCI. Which TWO components are essential for the retrieval step?

Select 2 answers

A.Fine-tuned generation model

B.Embedding model to convert chunks into vectors

C.Document chunking

D.Beam search decoder

E.Human feedback loop

AnswersB, C

Embeddings are required for vector search.

Why this answer

Chunking splits documents into manageable pieces, and embedding converts them into vectors for similarity search.

Full explanation →

355

MCQmedium

You are a data scientist at a legal firm. The firm uses OCR to digitize court documents and then indexes them in OCI OpenSearch for a RAG application. The application uses OCI Generative AI Service (Cohere Command) to answer questions about case law. Recently, the team noticed that the answers are often factually incorrect or include information not present in the retrieved documents. After reviewing the pipeline, you find that the chunking strategy splits documents into 512-token chunks with 128-token overlap. The embedding model is Cohere Embed v3 (English), and the retrieval returns the top 5 chunks. The LLM has a context window of 4096 tokens. The team suspects that the chunking strategy is causing loss of context. What is the best course of action to improve answer accuracy?

A.Increase the chunk size to 1024 tokens and overlap to 256 tokens.

B.Reduce the chunk overlap to 64 tokens to avoid redundancy.

C.Switch to a smaller LLM with a larger context window.

D.Increase the number of retrieved chunks from 5 to 10.

AnswerA

Larger chunks with more overlap preserve context better.

Why this answer

Increasing the chunk size to 1024 tokens and overlap to 256 tokens directly addresses the loss of context by ensuring each chunk contains more complete semantic units (e.g., entire paragraphs or legal arguments) while the larger overlap preserves continuity across chunk boundaries. This improves the quality of the embeddings and the relevance of retrieved chunks, leading to more factually accurate answers from the LLM.

Exam trap

The trap here is that candidates may assume increasing retrieval count (Option D) always improves accuracy, but in RAG systems, more chunks often introduce noise and dilute relevant context, whereas fixing the chunking strategy directly addresses the root cause of context loss.

How to eliminate wrong answers

Option B is wrong because reducing the overlap to 64 tokens would further fragment context, increasing the risk of missing critical information at chunk boundaries and worsening the factual inaccuracies. Option C is wrong because switching to a smaller LLM with a larger context window does not fix the root cause—poor chunking—and a smaller model may have lower reasoning capability, potentially degrading answer quality. Option D is wrong because increasing the number of retrieved chunks from 5 to 10 would introduce more noise and irrelevant content into the LLM's context, likely amplifying hallucinations rather than improving accuracy.

Full explanation →

356

MCQmedium

A data scientist needs to create vector embeddings for a multilingual customer feedback dataset to perform clustering analysis. Which OCI Generative AI embedding model should they choose?

A.Cohere Command R+

B.Cohere embed-english-v3.0

C.Meta Llama 3 70B

D.Cohere embed-multilingual-v3.0

AnswerD

This embedding model supports over 100 languages, suitable for multilingual clustering.

Why this answer

The embed-multilingual-v3.0 model is designed for multilingual content, making it the best choice for a multilingual dataset. The other options are either English-only or not embedding models.

Full explanation →

357

MCQhard

An OCI user observes that their Mistral model produces very repetitive text when temperature is set to 0.9 and top-p to 1.0. Which adjustment is most likely to reduce repetition?

A.Decrease top-p to 0.5

B.Increase temperature to 1.5

C.Enable frequency penalty

D.Set top-k to 1

AnswerC

Frequency penalty penalizes tokens based on their frequency in the generated text.

Why this answer

Enabling frequency penalty directly reduces the likelihood of the model repeating the same tokens by subtracting a penalty proportional to the token's existing frequency in the generated text. This is the most targeted way to combat repetitive output without altering the core sampling parameters (temperature and top-p) that control randomness.

Exam trap

A common mistake in Oracle OCI GenAI is assuming that reducing randomness (lower top-p or top-k) will fix repetition, when it often makes it worse, and that increasing temperature helps, but it risks incoherence. Frequency penalty is the targeted solution.

How to eliminate wrong answers

Option A is wrong because decreasing top-p to 0.5 narrows the set of tokens considered for sampling, which can actually increase repetition by focusing on the most probable tokens. Option B is wrong because increasing temperature to 1.5 makes the output more random and chaotic, which may reduce repetition but often at the cost of coherence and relevance. Option D is wrong because setting top-k to 1 forces greedy decoding (always picking the single most probable token), which dramatically increases repetition and is the opposite of what is needed.

Full explanation →

358

MCQeasy

A developer wants to interactively test different prompts and parameters (temperature, top_p, frequency_penalty) with a Cohere Command R model before integrating it into an application. Which tool should they use?

A.OCI Console Resource Manager

B.OCI SDK for Python

C.OCI Data Science Notebooks

D.OCI Generative AI Playground

AnswerD

The Playground allows interactive testing of models with adjustable parameters.

Why this answer

The OCI Generative AI Playground provides a no-code interface to test models, adjust parameters, and see real-time outputs. It is designed for experimentation before production use.

Full explanation →

359

MCQeasy

Your team has deployed a fine-tuned GPT-2 model on OCI Model Deployment for a simple text generation API. The model performs text completion for short prompts (e.g., 50 tokens). The endpoint is working but response times are over 10 seconds for these short prompts. The model size is approximately 500MB and you used a VM.Standard.E3.Flex shape (2 OCPU, 16GB RAM). The deployment is in a single replica with no autoscaling. You have verified that the network latency is minimal (<5ms). The model was trained in OCI Data Science using a GPU shape, but during deployment you selected a CPU shape to reduce cost. The model is a transformer-based neural network. You've also confirmed that the deployment is healthy and there are no errors in the logs. The memory usage is within limits. What is the most likely cause of the high latency?

A.High network latency between the client and the model endpoint

B.Model is too large for the VM.Standard.E3.Flex shape

C.Insufficient CPU resources for the model size

D.Missing GPU acceleration for inference

AnswerD

GPU acceleration is essential for fast inference on neural network models like GPT-2.

Why this answer

GPT-2 is a transformer-based neural network that relies heavily on matrix multiplications, which are far more efficiently executed on GPUs due to their parallel architecture. Even though the model is only 500MB, CPU inference for transformer models is notoriously slow because CPUs process sequential operations, while GPUs can parallelize the attention mechanism and feed-forward layers. The 10-second latency for a 50-token prompt is a classic symptom of missing GPU acceleration, as the CPU shape (2 OCPU) lacks the specialized tensor cores needed for fast transformer inference.

Exam trap

The trap here is that candidates might assume a 500MB model is 'small enough' for CPU inference, overlooking that transformer architecture—not model size—is the primary driver of latency, and that GPU acceleration is essential even for moderately sized transformer models.

How to eliminate wrong answers

Option A is wrong because the problem states that network latency is minimal (<5ms), so high network latency is not the cause. Option B is wrong because the model size of 500MB fits comfortably within the 16GB RAM of the VM.Standard.E3.Flex shape, and memory usage is confirmed to be within limits. Option C is wrong because while CPU resources are limited (2 OCPU), the core issue is not insufficient CPU resources per se, but rather that CPUs are architecturally unsuited for the parallel computations required by transformer models; even with more CPU cores, inference would still be significantly slower than with a GPU.

Full explanation →

360

MCQmedium

An OCI CLI command above returns embeddings for the phrase 'Hello world'. The developer notices that the embedding vector length is 384 dimensions. However, they expected 768 dimensions. What is the most likely cause?

A.The input text 'Hello world' is too short, causing dimension reduction.

B.The CLI result is truncated in the display.

C.The model 'cohere.embed-multilingual-light-v3.0' outputs 384-dimensional vectors.

D.The --truncate END flag reduces the dimension.

AnswerC

This specific model produces 384 dimensions; the 'light' version is smaller.

Why this answer

The Cohere model 'cohere.embed-multilingual-light-v3.0' is specifically designed to output 384-dimensional embeddings. The developer's expectation of 768 dimensions likely stems from familiarity with larger models like 'cohere.embed-english-v3.0', which outputs 1024 dimensions, or other models that produce 768-dimensional vectors. The embedding dimension is a fixed property of the model, not influenced by input length or CLI display settings.

Exam trap

Oracle often tests the misconception that embedding dimension is dynamically determined by input length or CLI flags, when in fact it is a static property of the chosen model.

How to eliminate wrong answers

Option A is wrong because embedding dimension is a fixed property of the model and does not change based on input text length; short inputs do not cause dimension reduction. Option B is wrong because the CLI does not truncate the embedding vector in the display; the full vector is returned, and any display truncation would be cosmetic and not affect the actual dimension count. Option D is wrong because the --truncate END flag controls how the input text is truncated to fit the model's token limit, not the output embedding dimension.

Full explanation →

361

MCQeasy

A user sends an inference request with the JSON parameters shown. They notice the model is returning very short responses. What is the most likely cause?

A.maxTokens is set too high

B.topP is set too high

C.The modelId is incorrect

D.temperature is set too low

AnswerD

Low temperature reduces randomness, often leading to shorter, safer outputs.

Why this answer

A low temperature value (close to 0) makes the model highly deterministic, reducing randomness and often leading to shorter, more conservative responses. In generative AI, temperature controls the probability distribution over tokens; lower values cause the model to favor the most likely tokens, which can result in repetitive or truncated outputs. The user's inference request likely includes a temperature setting that is too low, causing the model to produce very short responses.

Exam trap

Oracle often tests the misconception that maxTokens or topP control response length directly, when in fact temperature has a more subtle effect on output length by influencing token diversity and repetition.

How to eliminate wrong answers

Option A is wrong because maxTokens sets the maximum number of tokens the model can generate; setting it too high would allow longer responses, not shorter ones. Option B is wrong because topP (nucleus sampling) controls the cumulative probability threshold for token selection; setting it too high would include more diverse tokens, potentially leading to longer or more varied responses, not shorter ones. Option C is wrong because an incorrect modelId would typically cause an error or unexpected behavior (e.g., model not found), not consistently produce very short responses.

Full explanation →

362

MCQmedium

A company is fine-tuning a Llama model on OCI with dedicated AI cluster. They want to use their own training data stored in Oracle Object Storage. What must they do to ensure the fine-tuning job can access the data?

A.Upload the data to the dedicated cluster's local storage.

B.Use OCI Data Flow to transfer data.

C.Configure a resource principal for the cluster to read the bucket.

D.Create a pre-authenticated request for the bucket.

AnswerC

A service principal grants the cluster permissions to access resources like Object Storage.

Why this answer

OCI Generative AI dedicated AI clusters require a resource principal (an IAM principal) to authenticate and authorize access to data stored in Oracle Object Storage. By configuring the cluster with a resource principal that has read permissions on the bucket, the fine-tuning job can securely read the training data without exposing credentials or using public URLs.

Exam trap

The trap here is that candidates confuse OCI pre-authenticated requests (PARs) with secure IAM-based access, not realizing that PARs are designed for temporary, user-facing sharing and lack the auditability and policy control required for automated ML workloads on OCI.

How to eliminate wrong answers

Option A is wrong because dedicated AI clusters do not provide local storage for user data; training data must remain in Object Storage and be accessed remotely. Option B is wrong because OCI Data Flow is a separate Spark-based service for data processing, not a mechanism for granting access to Object Storage from a Generative AI cluster. Option D is wrong because a pre-authenticated request (PAR) provides temporary public access to a bucket or object, which is insecure and not intended for automated, long-running fine-tuning jobs that require IAM-based access control.

Full explanation →

363

MCQhard

An LLM generates a response that contains a plausible-sounding but factually incorrect statement about a historical event. This is an example of which known limitation?

A.Knowledge cutoff

B.Hallucination

C.Bias in training data

D.Context length constraint

AnswerB

Hallucination is when the model produces factually incorrect or nonsensical content that appears plausible.

Why this answer

Hallucination in LLMs refers to the generation of content that is plausible-sounding but factually incorrect or nonsensical. This occurs when the model's probabilistic next-token prediction produces statements that are not grounded in its training data or real-world facts, often due to overgeneralization or lack of factual recall mechanisms.

Exam trap

A common pitfall is confusing hallucination with knowledge cutoff. Hallucination generates incorrect information even on topics within the training data's timeframe, while knowledge cutoff only explains missing recent events. This question tests the ability to distinguish these two limitations.

How to eliminate wrong answers

Option A is wrong because knowledge cutoff refers to the date after which the model has no training data, not to the generation of incorrect facts within its training period. Option C is wrong because bias in training data leads to systematic skewing of outputs (e.g., stereotypes), not to isolated factually incorrect statements about specific events. Option D is wrong because context length constraint limits the amount of input text the model can process at once, not the factual accuracy of its generated responses.

Full explanation →

364

MCQhard

A development team wants to generate code snippets from natural language. Which model strategy should they adopt?

A.Use a code-specific model like Code Llama.

B.Use a general-purpose LLM like Llama 2.

C.Use a multimodal model.

D.Use an embedding model for text.

AnswerA

Correct: Code-specific models are fine-tuned for code generation.

Why this answer

Code Llama is a specialized variant of Llama 2 that has been fine-tuned on code datasets, enabling it to generate syntactically and semantically correct code from natural language prompts. This makes it the optimal choice for code generation tasks, as general-purpose LLMs lack the targeted training on code structures and programming languages.

Exam trap

Oracle often tests the distinction between general-purpose and domain-specific models, and the trap here is that candidates assume any large language model can handle code generation equally well, overlooking the critical fine-tuning on code corpora that makes Code Llama superior for this task.

How to eliminate wrong answers

Option B is wrong because a general-purpose LLM like Llama 2 is trained on diverse text but not specifically optimized for code, leading to higher rates of syntax errors and logical inconsistencies in generated code. Option C is wrong because multimodal models process images, audio, and text, but code generation from natural language does not require multiple modalities and would add unnecessary complexity without improving code quality. Option D is wrong because embedding models are designed to convert text into vector representations for similarity search or clustering, not for generating new text or code snippets.

Full explanation →

365

Multi-Selecthard

A practitioner is choosing a model for a code generation assistant that must run on OCI with low latency. Which THREE considerations are most important?

Select 3 answers

A.Model size (number of parameters)

B.Whether the model is available as a managed endpoint or requires self-hosting on OCI GPU shapes

C.The cost of fine-tuning the model on internal codebases

D.Availability of a model specifically fine-tuned for code (e.g., Code Llama, StarCoder)

E.The model's context window length

AnswersA, B, D

Smaller models generally have lower inference latency, which is critical for real-time code generation.

Why this answer

Smaller models typically have lower latency, model families (Llama, Code Llama) are designed for code, and deployment choices affect latency. Context window size and fine-tuning cost are secondary or not directly relevant to latency.

Full explanation →

366

Multi-Selecthard

Which THREE factors directly influence the quality of responses in a RAG system? (Choose three.)

Select 3 answers

A.The prompt template used to ask the LLM

B.The chunk size used during document processing

C.The temperature parameter of the LLM

D.The number of GPUs allocated to the LLM

E.The choice of embedding model

AnswersA, B, E

A well-structured prompt helps the LLM use the context properly.

Why this answer

The prompt template directly controls how the LLM interprets the retrieved context and formulates its response. A well-structured prompt with clear instructions, context formatting, and output constraints significantly improves response relevance and accuracy, while a poorly designed prompt can lead to hallucinations or off-topic answers.

Exam trap

Oracle certification exams often test the distinction between factors that directly influence response quality (prompt, chunk size, embedding model) versus factors that affect performance or output style (temperature, GPU count), leading candidates to mistakenly select temperature or hardware options.

Full explanation →

367

MCQmedium

Refer to the exhibit. A developer ran the OCI CLI command shown and received the JSON output. What does the output indicate about the model's confidence and why?

A.The model is uncertain because all scores are roughly equal.

B.The model is neutral because the neutral score is lowest.

C.The model is unsure because the scores are probabilities that sum to 1.

D.The model is highly confident the text is positive, as indicated by the 0.98 score.

AnswerD

0.98 is very close to 1, indicating high confidence.

Why this answer

The JSON output shows a sentiment score of 0.98 for 'positive', which is very close to 1.0, indicating the model is highly confident that the text is positive. In sentiment analysis models, scores represent probabilities for each class, and a value near 1.0 for one class with much lower scores for others reflects strong confidence.

Exam trap

Oracle often tests the distinction between the sum of probabilities equaling 1 (a mathematical property) and the actual confidence level indicated by the distribution of those probabilities, leading candidates to mistakenly choose option C.

How to eliminate wrong answers

Option A is wrong because the scores are not roughly equal; the positive score (0.98) is significantly higher than the negative (0.01) and neutral (0.01) scores, indicating high confidence, not uncertainty. Option B is wrong because the neutral score being lowest does not imply neutrality; the model is highly confident the text is positive, not neutral. Option C is wrong because while the scores do sum to 1 (as probabilities should), this fact alone does not indicate uncertainty; the distribution of probabilities matters, and here the high positive score shows confidence.

Full explanation →

368

MCQeasy

A developer is using OCI Data Science to create a RAG pipeline. They have ingested documents into a vector store using OCI Generative AI's text-embedding model. During testing, they notice that queries return very few results (often 0 or 1) even when the knowledge base contains relevant documents. They have set the top-k parameter to 10. What is the most likely cause?

A.The similarity threshold is set too high, filtering out most results.

B.The documents were chunked with too small a chunk size, losing key information.

C.The embedding model's dimensionality is too low to capture semantic differences.

D.The vector search index is not configured with the correct distance metric.

AnswerA

A high similarity threshold filters out many results, causing very few to pass.

Why this answer

A high similarity threshold (e.g., >0.9) can exclude many relevant results, leading to few or zero results even with top-k set to 10. Option B: chunk size affects the granularity of text but not directly the number of results returned; small chunks can still be retrieved if the threshold is appropriate. Option C: dimensionality of the embedding model is fixed and does not directly cause zero results; low dimensionality may reduce semantic precision but not eliminate results.

Option D: distance metric affects how similarity is computed but not the count; the index can still return results ranked by the chosen metric.

Full explanation →

369

MCQhard

An application mixes RAG with other data sources. The vector search returns too many irrelevant chunks. What is the best approach to filter them?

A.Use a reranker model

B.Use exact search instead of ANN

C.Reduce the number of retrieved chunks

D.Increase chunk size

AnswerA

A reranker scores retrieved chunks by relevance, filtering out irrelevant ones.

Why this answer

A reranker model (Option A) is the best approach because it takes the initial set of retrieved chunks and re-orders them based on semantic relevance to the query, effectively filtering out irrelevant chunks. Unlike simple vector similarity, a reranker uses cross-encoding to evaluate the query-chunk pair as a whole, which significantly improves precision when mixing RAG with other data sources.

Exam trap

Oracle often tests the misconception that reducing the number of retrieved chunks (Option C) is a valid filter, but the trap is that this only limits output size without improving relevance—reranking is the correct technique to reorder and discard irrelevant results.

How to eliminate wrong answers

Option B is wrong because exact search (e.g., brute-force k-NN) retrieves the same chunks as ANN but without approximation; it does not filter irrelevant chunks—it only guarantees the true nearest neighbors, which may still be irrelevant if the vector representation is poor. Option C is wrong because reducing the number of retrieved chunks (e.g., lowering top_k) risks missing relevant chunks and does not address the core problem of irrelevant chunks being ranked too high. Option D is wrong because increasing chunk size makes each chunk more likely to contain irrelevant content, potentially worsening the problem by diluting relevant information with noise.

Full explanation →

370

MCQhard

An OCI user is comparing two embedding models: one with 768 dimensions and another with 1024 dimensions. Which of the following trade-offs is most relevant?

A.The 1024-dimensional model always yields better accuracy with no additional cost

B.The 768-dimensional model is always faster and more storage-efficient

C.Higher dimensions may capture more fine-grained semantic information but require more storage and slower similarity search

D.Dimension size has no impact on performance or storage

AnswerC

This is the standard trade-off.

Why this answer

Higher dimensions can capture more nuanced semantics but require more storage and may increase retrieval latency.

Full explanation →

371

Multi-Selectmedium

A prompt engineer is designing a prompt for legal document analysis. The model should output JSON with a specific structure. Which TWO practices are essential to include in the prompt?

Select 2 answers

A.Define the JSON schema in the system prompt

B.Provide a concrete example of the desired JSON output

C.Use a stop sequence of "}"

D.Increase max tokens to 5000

E.Set temperature to 1.0 for creativity

AnswersA, B

Defining the schema sets the expected structure.

Why this answer

Defining the JSON schema in the system prompt explicitly instructs the model on the exact structure and data types required for the output. This is a foundational practice in prompt engineering for structured data generation, as it reduces ambiguity and ensures the model adheres to a predefined format, which is critical for downstream parsing in legal document analysis.

Exam trap

The trap here is that candidates confuse 'stop sequences' (used to truncate output) with 'format constraints' (like JSON schemas), or assume higher token limits and creative temperature settings are universally beneficial, when in fact they degrade reliability for structured data tasks.

Full explanation →

372

MCQeasy

A developer wants to implement a simple RAG pipeline using OCI Language's text generation and embedding models. Which OCI SDK method is used to generate embeddings for a text chunk?

A.embed_text

B.generate_embeddings

C.encode_text

D.create_embedding

AnswerA

`embed_text` is the correct method to call for generating embeddings from text.

Why this answer

The correct OCI SDK method for generating embeddings for a text chunk is `embed_text`. This method is part of the OCI Language service's `AIServiceLanguageClient` and directly returns vector representations of input text, which are essential for RAG pipelines to enable semantic search and retrieval.

Exam trap

The trap here is that candidates confuse OCI SDK method names with those from other cloud providers (e.g., OpenAI's `create_embedding` or generic `encode_text`), leading them to select a plausible-sounding but incorrect option.

How to eliminate wrong answers

Option B is wrong because `generate_embeddings` is not a valid method in the OCI Language SDK; the correct method name is `embed_text`. Option C is wrong because `encode_text` is not an OCI SDK method; it resembles a generic function name from other frameworks (e.g., Hugging Face) but does not exist in OCI's API. Option D is wrong because `create_embedding` is not an OCI SDK method; it is a method name used by OpenAI's API, not by OCI Language.

Full explanation →

373

MCQmedium

A data scientist is using OCI Generative AI to process a large batch of legal documents. The total cost is higher than expected. Which factor is most likely the primary driver of cost?

A.The number of API requests made

B.The total number of tokens processed (input + output)

C.The choice of sampling strategy (e.g., top-k vs greedy)

D.The latency of the inference endpoint

AnswerB

Pricing is typically per token; longer documents mean more tokens and higher cost.

Why this answer

In OCI Generative AI, pricing is primarily based on the total number of tokens processed, which includes both input (prompt) and output (generated) tokens. Processing large batches of legal documents generates high token counts due to lengthy text inputs and verbose outputs, directly increasing cost. The number of API requests alone does not determine cost—a single request with many tokens costs more than many requests with few tokens.

Exam trap

The 1Z0-1127 exam often tests the misconception that API request count is the primary cost driver, when in reality token-based pricing means a single large request can cost more than hundreds of tiny requests.

How to eliminate wrong answers

Option A is wrong because OCI Generative AI charges per token, not per API request; a single request with a large prompt and long response incurs higher cost than many small requests. Option C is wrong because sampling strategy (e.g., top-k vs greedy) affects output diversity and quality, not the token count or pricing model. Option D is wrong because latency of the inference endpoint impacts response time and throughput, not the cost per token or total cost.

Full explanation →

374

MCQeasy

Which API should a developer use to send a multi-turn conversation history to an LLM, including a system message and previous user/assistant exchanges, using OCI Generative AI?

A.Summarisation API

B.Generate API

C.Chat API

D.Embedding API

AnswerC

Chat API supports system prompts and conversation history.

Why this answer

The Chat API is designed for multi-turn conversations with system prompts and history management.

Full explanation →

375

MCQhard

A developer is using the Cohere Command R model with a preamble set to "You are a helpful assistant." The user message contains a question. The model's response is factual but lacks detail. To encourage more comprehensive answers, which modification should be made?

A.Use a stop sequence like "\n" to force longer responses

B.Set the temperature to 0.0

C.Add "Provide a thorough and detailed response." to the preamble

D.Increase the max_tokens parameter to 1000

AnswerC

Preamble is the system message; instructing the model to be detailed influences its behavior.

Why this answer

Adding a directive in the preamble, such as "Provide thorough and detailed responses," sets the desired behavior from the start.

Full explanation →

Oracle Cloud Infrastructure Generative AI Professional 1Z0-1127-25 (1Z0-1127-25) — Questions 301–375