Courseiva

Oracle Cloud Infrastructure Generative AI Professional 1Z0-1127-25 (1Z0-1127-25) — Questions 226300

768 questions total · 11pages · All types, answers revealed

Page 3

Page 4 of 11

Page 5
226
Multi-Selecthard

An organization is deploying a RAG application with Oracle AI Vector Search. They need to ensure that the vector index supports low-latency queries and can handle updates to the underlying documents (inserts, deletes, modifications) without significant performance degradation. Which two index features should they consider? (Choose TWO.)

Select 2 answers
A.Use a VECTOR data type with a default B-tree index
B.Use an IVF index with periodic rebuilds to maintain performance after many updates
C.Enable exact nearest neighbor search to avoid index maintenance
D.Disable indexing and rely on full table scan for simplicity
E.Use an HNSW index, which supports incremental updates and provides low-latency search
AnswersB, E

IVF indexes can be rebuilt periodically to handle updates; with proper maintenance, they can provide low-latency queries.

Why this answer

An IVF (Inverted File) index with periodic rebuilds is well-suited for RAG applications that experience frequent updates (inserts, deletes, modifications). IVF indexes are designed for approximate nearest neighbor search, offering low-latency queries, but they can degrade over time as data changes; periodic rebuilds restore performance without requiring a full re-index of the entire dataset. This approach balances query speed with update tolerance, making it a practical choice for dynamic document collections in Oracle AI Vector Search.

Exam trap

The 1Z0-1127 exam often tests the misconception that HNSW indexes are always superior for dynamic workloads, but the question explicitly asks for two features that support low-latency queries and handle updates, and both IVF with periodic rebuilds and HNSW are valid; the trap is that candidates might overlook the periodic rebuild requirement for IVF or incorrectly assume HNSW is the only option, leading them to select only one correct answer or to dismiss IVF entirely.

227
MCQeasy

Which OCI Generative AI service component is designed to convert text into vector representations for use in semantic search?

A.Chat API
B.Summarisation API
C.Generate API
D.Embedding API
AnswerD

Embedding API produces vector embeddings from text inputs.

Why this answer

The Embedding API generates vector embeddings from text, which can be used for semantic search, clustering, etc.

228
MCQhard

A developer is using the Cohere Command model via OCI Generative AI and wants to ground responses in a specific uploaded document. Which syntax should be used in the preamble to enable document-grounded generation?

A.Set a system message with 'DOCUMENT: ...'
B.Include the document in the user message with 'Document: ...'
C.Pass the document as a separate parameter named 'context'
D.Use the preamble field with '<doc>document text</doc>'
AnswerD

Cohere's API expects the document to be embedded in the preamble using <doc> tags for grounded generation.

Why this answer

Cohere's document-grounded generation uses a special preamble format that includes document references. The correct syntax is to enclose document text in <doc> tags within the preamble.

229
Multi-Selectmedium

A company is designing a generative AI application using OCI Generative AI. Which two factors should be considered when selecting the appropriate model? (Choose two.)

Select 2 answers
A.Model's training data cutoff date
B.Availability in all OCI regions
C.Supported languages
D.Maximum token output limit
E.Built-in safety filters
AnswersA, D

The cutoff date indicates how recent the model's knowledge is.

Why this answer

The model's training data cutoff date determines the temporal scope of the model's knowledge. For generative AI applications requiring up-to-date information or compliance with data recency requirements, selecting a model with a cutoff date that aligns with the use case is critical. OCI Generative AI models have specific cutoff dates (e.g., June 2023 for certain models), and using a model with an older cutoff may produce outdated or factually incorrect responses.

Exam trap

The trap here is that candidates often confuse service-level features (like safety filters or regional availability) with model-specific selection criteria, leading them to pick options that are technically true but irrelevant to the core decision of choosing the right model for a generative AI application.

230
MCQmedium

A retail company uses OCI Generative AI Agents to power a product recommendation chatbot on their e-commerce website. The chatbot is integrated with a knowledge base containing product descriptions, customer reviews, and inventory data. Recently, the chatbot has started recommending out-of-stock products frequently, leading to customer frustration. The development team verified that the knowledge base is updated in real-time with inventory data. The chatbot's configuration uses a chunking strategy with a chunk size of 500 tokens and an overlap of 50 tokens. The team suspects the issue is related to how the agent retrieves information. They have access to OCI Logging and Monitoring. Which course of action should the team take first?

A.Decrease the chunk size to 250 tokens to make chunks more specific.
B.Reduce the temperature parameter of the model to 0.2 to reduce hallucinations.
C.Enable auto-scaling on the AI cluster to improve response speed.
D.Increase the chunk overlap from 50 to 150 tokens to ensure inventory status is captured in multiple chunks.
AnswerD

Greater overlap ensures that inventory updates are not missed, improving the relevance of retrieved context.

Why this answer

The core issue is that the chatbot retrieves chunks that contain product descriptions but may miss the inventory status because the chunking strategy does not reliably include both pieces of information together. Increasing the chunk overlap from 50 to 150 tokens ensures that inventory data, which may be at the boundary of a chunk, is captured in multiple overlapping chunks, thereby increasing the likelihood that the retrieval step returns a chunk containing both the product and its current stock level. This directly addresses the retrieval gap without altering model behavior or infrastructure.

Exam trap

Oracle often tests the misconception that retrieval issues are always solved by adjusting model parameters (like temperature) or infrastructure scaling, when the real fix lies in tuning the chunking strategy to ensure critical metadata is not lost at chunk boundaries.

How to eliminate wrong answers

Option A is wrong because decreasing chunk size to 250 tokens would make chunks more specific but would also increase the number of chunks and the risk that inventory status is split across even more chunks, potentially worsening the problem. Option B is wrong because reducing the temperature parameter reduces randomness in generation but does not affect how the agent retrieves information from the knowledge base; the issue is retrieval, not hallucination. Option C is wrong because enabling auto-scaling improves response speed and throughput but does not change the content or structure of the chunks being retrieved, so it cannot fix the missing inventory data.

231
MCQmedium

A company wants to use OCI Generative AI to summarize customer support tickets. They need to ensure that the model does not output any sensitive information. Which technique should they implement?

A.Prompt engineering to instruct the model to exclude sensitive information.
B.Use a smaller model that is less likely to memorize data.
C.Enable content filtering on the endpoint.
D.Disable the use of training data in the endpoint configuration.
AnswerA

Carefully crafted prompts can guide the model to avoid leaking sensitive data.

Why this answer

Prompt engineering is the correct technique because it allows the company to explicitly instruct the generative AI model to exclude sensitive information from its outputs. By crafting a system prompt or user prompt with specific directives (e.g., 'Do not include any personally identifiable information, account numbers, or confidential data in your summary'), the model's behavior is directly controlled at inference time. This is a lightweight, flexible approach that does not require changing the model architecture or endpoint configuration, and it is the most direct way to enforce output constraints in OCI Generative AI.

Exam trap

Oracle often tests the misconception that disabling training data or using a smaller model can prevent sensitive output, when in fact prompt engineering is the primary technique for controlling model behavior at inference time in OCI Generative AI.

How to eliminate wrong answers

Option B is wrong because using a smaller model does not guarantee the exclusion of sensitive information; smaller models can still memorize and output sensitive data from their training set, and model size is unrelated to output filtering. Option C is wrong because content filtering on the endpoint typically blocks predefined categories (e.g., hate speech, violence) but is not designed to dynamically detect and remove sensitive business data like customer support ticket details. Option D is wrong because disabling the use of training data in the endpoint configuration (e.g., setting 'trainingDataConsent' to false) only prevents the model from being fine-tuned or retrained on the input data; it does not affect the model's output behavior during inference, so sensitive information can still appear in summaries.

232
MCQmedium

A developer is getting a 401 Unauthorized error when calling the OCI Generative AI inference API. What is the most likely cause?

A.The API endpoint has reached its rate limit
B.The request is missing or has an invalid authentication signature
C.The model does not support the requested parameters
D.The model is not deployed
AnswerB

401 Unauthorized specifically indicates authentication failure.

Why this answer

A 401 Unauthorized error specifically indicates a failure in authentication, not authorization or resource availability. The OCI Generative AI inference API requires every request to include a valid signature based on the OCI Signature Version 1 algorithm (RFC 2104 HMAC-SHA256). If the request is missing the Authorization header or the signature is malformed (e.g., incorrect key ID, mismatched signing string, or expired timestamp), the API gateway rejects it with a 401 response.

Exam trap

Oracle often tests the distinction between HTTP status codes (401 vs 403 vs 429 vs 404) to see if candidates confuse authentication failures with authorization, rate limiting, or resource availability issues.

How to eliminate wrong answers

Option A is wrong because a rate limit exceeded (HTTP 429) would return a 'Too Many Requests' error, not a 401 Unauthorized. Option C is wrong because unsupported parameters typically result in a 400 Bad Request error, not a 401. Option D is wrong because a model that is not deployed would return a 404 Not Found or a 400 error, as the endpoint itself would be unreachable or the model ID invalid, not an authentication failure.

233
MCQmedium

A developer is using the OCI GenAI Chat API to build a multi-turn customer support chatbot. They want the assistant to always introduce itself as 'SupportBot' and never mention being an AI. How should they configure the API call?

A.Use the Generate API instead and include the instruction in the prompt
B.Set the preamble override to a JSON object with the assistant's identity
C.Set the system message (preamble) to 'You are SupportBot. You must never mention that you are an AI.'
D.Set the first user message to 'Introduce yourself as SupportBot and never say you are an AI.'
AnswerC

The system message (preamble) defines the assistant's persona and instructions.

Why this answer

The OCI GenAI Chat API supports a 'system message' or 'preamble' parameter that sets the assistant's behavior and identity for the entire conversation. By setting the preamble to 'You are SupportBot. You must never mention that you are an AI.', the developer enforces the desired persona and restriction across all turns, ensuring the assistant introduces itself as SupportBot and avoids any reference to being an AI.

Exam trap

A common trap in the OCI GenAI exam is confusing the system message (preamble) parameter with a regular user message. Candidates may think placing instructions in the first user message is sufficient, but only the dedicated preamble parameter enforces behavior across all turns. The Chat API requires explicit use of the preamble to set persistent assistant identity and restrictions.

How to eliminate wrong answers

Option A is wrong because the Generate API is a single-turn text generation endpoint that does not support multi-turn conversation context or persistent system instructions; using it would require manually managing conversation history and re-injecting the instruction in every prompt, which is inefficient and error-prone. Option B is wrong because the OCI GenAI Chat API does not accept a 'preamble override' as a JSON object; the correct parameter is a plain text string for the system message (preamble), and a JSON object would be rejected or misinterpreted. Option D is wrong because setting the first user message to include the instruction does not persist across turns; the assistant might follow it initially but could deviate in subsequent responses, and the instruction is not enforced as a system-level constraint.

234
MCQmedium

A company wants to use OCI Generative AI Agents to build a question-answering system over documents stored in OCI Object Storage. Which component acts as the knowledge source for the agent?

A.A Dedicated AI Cluster
B.The OCI Generative AI Playground
C.An endpoint created via the InferenceClient
D.A knowledge base created from the Object Storage bucket
AnswerD

The knowledge base indexes documents from the data source (Object Storage) for retrieval.

Why this answer

In OCI Generative AI Agents, a knowledge base is the indexed representation of data sources (like Object Storage buckets). The agent uses the knowledge base to retrieve relevant information.

235
MCQeasy

You need to convert a set of customer support tickets into vector embeddings for a similarity search application. Which OCI Generative AI model should you use?

A.Cohere Rerank
B.Cohere Embed (e.g., embed-english-v3.0)
C.Meta Llama 3
D.Cohere Command R
AnswerB

Cohere Embed models are specifically designed to generate dense vector embeddings from text, ideal for similarity search and retrieval tasks.

Why this answer

The Cohere Embed models are designed for text-to-vector embedding. The other options are for text generation or reranking.

236
MCQhard

A prompt engineer is designing a system that generates SQL queries from natural language. The model sometimes produces unsafe queries (e.g., DROP TABLE). Which constraint in the system prompt would BEST mitigate this risk?

A.Use a few-shot prompt with only safe SELECT examples
B.Use role prompting: 'You are an expert SQL developer'
C.Set frequency penalty to 1.0 to avoid repetitive unsafe patterns
D.Include a constraint: 'You may only generate SELECT statements. Do not generate DDL or DML statements like DROP, DELETE, INSERT, or UPDATE.'
AnswerD

This explicit constraint directly addresses the safety concern.

Why this answer

It explicitly prohibits the model from generating DDL (Data Definition Language) and DML (Data Manipulation Language) statements, directly addressing the risk of unsafe queries like DROP TABLE. By constraining the output to only SELECT statements, the prompt enforces a strict policy that prevents the model from producing destructive or modifying SQL commands, which is the most effective mitigation among the options.

Exam trap

Oracle often tests the misconception that implicit guidance (like few-shot examples or role prompting) is sufficient to enforce safety, when in fact only explicit, unambiguous constraints can reliably prevent the model from generating prohibited outputs.

How to eliminate wrong answers

Option A is wrong because a few-shot prompt with only safe SELECT examples does not explicitly forbid unsafe queries; the model may still generalize from its training data and generate DROP or DELETE statements when faced with ambiguous or malicious input. Option B is wrong because role prompting ('You are an expert SQL developer') does not impose any behavioral constraint; an expert developer might still generate DDL or DML statements if the user asks for them, as the role does not inherently restrict output. Option C is wrong because setting a frequency penalty to 1.0 reduces the likelihood of repetitive patterns but does not prevent the model from generating unsafe queries; it only discourages token repetition, not the generation of specific dangerous commands.

237
MCQeasy

A data scientist needs to fine-tune a model on OCI Generative AI. Which of the following is a required parameter in the fine-tuning request?

A.hyperparameters
B.model_name
C.dataset_type
D.All of the above
AnswerD

All three (model_name, dataset_type, hyperparameters) are required for a fine-tuning request.

Why this answer

In OCI Generative AI, the fine-tuning request requires all three parameters: hyperparameters (to define training behavior like learning rate and epochs), model_name (to specify the base model being fine-tuned), and dataset_type (to indicate the format of the training data, such as 'TEXT' or 'MULTI_TURN'). Therefore, 'All of the above' is correct because each listed option is a mandatory field in the fine-tuning API call.

Exam trap

Oracle often tests the 'All of the above' pattern when each individual option is factually correct but candidates incorrectly assume only one is required, missing the comprehensive nature of the fine-tuning request.

How to eliminate wrong answers

Option A is wrong because hyperparameters are indeed required, but the question asks for 'a required parameter' and the correct answer includes all options, so selecting only A would be incomplete. Option B is wrong because model_name is required, but again, it is not the only required parameter. Option C is wrong because dataset_type is required, but the question expects the comprehensive answer that all three are necessary.

The trap is that each individual option is technically required, but the question is designed to test whether you know that all three are mandatory in the fine-tuning request.

238
Multi-Selecteasy

Which two are essential components of the Transformer architecture? (Select TWO)

Select 2 answers
A.Pooling layers
B.Recurrent connections
C.Self-attention mechanism
D.Feed-forward neural network
E.Convolutional layers
AnswersC, D

Correct: Core component of Transformers.

Why this answer

The self-attention mechanism is essential because it allows each token in the input sequence to attend to every other token, capturing long-range dependencies without the sequential bottleneck of RNNs. This mechanism computes attention scores using queries, keys, and values, enabling parallel processing and forming the core of the Transformer's ability to model context.

Exam trap

Oracle often tests the misconception that Transformers still use recurrence or convolution for sequence processing, when in fact they rely solely on self-attention and feed-forward networks.

239
MCQmedium

Refer to the exhibit. What is the solution?

A.Use a different base model that supports fine-tuning.
B.Change the learning rate.
C.Increase the training epochs.
D.Use a different compartment.
AnswerA

The error indicates the base model does not support fine-tuning; switch to a supported model.

Why this answer

The exhibit indicates that the base model does not support fine-tuning, which is a prerequisite for adapting a large language model to a specific task or domain. Using a different base model that supports fine-tuning allows the model to be customized through supervised learning on task-specific data, enabling it to learn new patterns and improve performance. This is the correct solution because without fine-tuning capability, the model cannot be effectively adapted regardless of other hyperparameter adjustments.

Exam trap

Oracle often tests the distinction between hyperparameter tuning (learning rate, epochs) and fundamental model capability (fine-tuning support), leading candidates to mistakenly choose a hyperparameter adjustment when the core issue is that the model cannot be fine-tuned at all.

How to eliminate wrong answers

Option B is wrong because changing the learning rate only affects the optimization process during training, but if the base model does not support fine-tuning, no amount of learning rate adjustment will enable the model to be trained on new data. Option C is wrong because increasing the training epochs will not help if the model cannot be fine-tuned at all; epochs only matter when the model is actually being trained or fine-tuned. Option D is wrong because using a different compartment (a tenancy or organizational boundary in Oracle Cloud Infrastructure) does not change the underlying model's architecture or its ability to be fine-tuned; it only affects resource isolation and access control.

240
MCQmedium

A regulatory compliance team needs to restrict access to the OCI Generative AI service so that only users in the 'AI_Engineers' group can create fine-tuning jobs and endpoints. Which IAM policy statement should be used?

A.Allow group AI_Engineers to inspect generative-ai-family in compartment ABC
B.Allow group AI_Engineers to manage generative-ai-family in tenancy
C.Allow group AI_Engineers to read generative-ai-family in compartment ABC
D.Allow group AI_Engineers to use generative-ai-family in compartment ABC
AnswerD

'use' permission includes the ability to create, update, and delete generative AI resources.

Why this answer

To allow a group to manage generative AI resources, you allow them to use the 'generative-ai-family' resource type. Specific verbs like 'inspect' limit visibility but not management. 'Read' only allows viewing, not creating.

241
MCQhard

A company deploys a large language model on a dedicated AI cluster with 4 nodes. The model requires 128 GB of memory per instance, but the nodes have only 64 GB each. During inference, the nodes experience out-of-memory errors. What is the best solution?

A.Enable model parallelism across nodes
B.Increase the number of nodes to 8
C.Upgrade to higher memory node shapes
D.Reduce the batch size in inference requests
AnswerA

Model parallelism distributes the model across nodes, enabling inference with the available memory.

Why this answer

Model parallelism splits the model's layers or parameters across multiple nodes, allowing the 128 GB model to be distributed across the 4 nodes (each with 64 GB) so that no single node exceeds its memory capacity. This is the best solution because it directly addresses the memory constraint without requiring additional hardware or sacrificing inference throughput, and it is a standard technique for deploying large language models on distributed AI clusters.

Exam trap

Oracle often tests the misconception that scaling out (more nodes) or scaling down (batch size) can fix memory constraints for large models, but the trap here is that the model's parameter memory is fixed and cannot be reduced by batch size changes, and adding more nodes without parallelism still leaves each node unable to host the full model.

How to eliminate wrong answers

Option B is wrong because increasing the number of nodes to 8 does not solve the fundamental issue: each node still has only 64 GB, and the model requires 128 GB per instance; without model parallelism, each node would still try to load the entire model and fail. Option C is wrong because upgrading to higher memory node shapes (e.g., 128 GB per node) would work but is often cost-prohibitive or unavailable, and the question asks for the best solution given the existing cluster; model parallelism is more efficient and scalable. Option D is wrong because reducing the batch size reduces per-request memory usage but does not reduce the model's parameter memory footprint (128 GB), so the model itself still cannot fit into a single node's 64 GB memory.

242
MCQmedium

A company uses OCI Generative AI Service to build a chatbot for customer support. They notice that the model sometimes generates inappropriate responses. What is the MOST effective way to mitigate this without retraining the model?

A.Fine-tune the model with curated safe examples
B.Configure system instructions to define acceptable behavior
C.Reduce the temperature parameter to 0
D.Use the moderation API to filter responses
AnswerB

System instructions constrain the model's output at inference time without retraining.

Why this answer

Configuring system instructions is the most effective approach because it allows you to define the model's behavior and constraints at inference time without modifying the underlying model weights. In OCI Generative AI Service, system instructions act as a persistent prompt that guides the model's responses, enabling you to explicitly prohibit inappropriate content and enforce safety guidelines. This is a non-invasive, immediate mitigation that does not require the time, cost, or data preparation associated with retraining or fine-tuning.

Exam trap

Oracle often tests the distinction between inference-time controls (like system instructions) and training-time modifications (like fine-tuning), trapping candidates who assume that only retraining can fix behavioral issues, when in fact prompt-level constraints are the fastest and most practical solution for immediate mitigation.

How to eliminate wrong answers

Option A is wrong because fine-tuning requires retraining the model with curated datasets, which is time-consuming, resource-intensive, and contradicts the question's constraint of 'without retraining the model.' Option C is wrong because reducing the temperature to 0 makes the model deterministic and less creative, but it does not prevent inappropriate responses—it only reduces randomness, not the likelihood of generating harmful content based on learned patterns. Option D is wrong because OCI Generative AI Service does not have a built-in 'moderation API' like some other cloud providers; while you could implement a separate content filter, this would be an external post-processing step rather than a direct configuration of the model's behavior, and the question asks for the most effective method within the service itself.

243
MCQmedium

A team is deploying a generative AI model using OCI Functions for serverless inference. They are experiencing cold start latency of over 10 seconds for the first invocation after idle periods. What is the best strategy to reduce cold start latency?

A.Migrate the inference to OCI Data Flow for better performance.
B.Use provisioned concurrency to keep a set number of function instances warm.
C.Reduce the function timeout to force faster execution.
D.Increase the memory allocation for the function.
AnswerB

Provisioned concurrency eliminates cold start by pre-warming instances.

Why this answer

OCI Functions supports provisioned concurrency, which keeps a specified number of instances warm. Option A (increasing memory) can reduce cold start but not as effectively. Option C (reducing timeout) might cause failures.

Option D (using OCI Data Flow) is for data processing, not inference.

244
Multi-Selecteasy

Which TWO are best practices for securing a generative AI endpoint on OCI? (Select TWO)

Select 2 answers
A.Enable OCI Logging for audit
B.Use a public endpoint with IP restrictions
C.Disable authentication for internal use
D.Store API keys in OCI Vault
E.Use a dedicated AI cluster with a private subnet
AnswersD, E

OCI Vault securely manages secrets and API keys.

Why this answer

OCI Vault provides a secure, centralized service for storing and managing API keys used to authenticate requests to generative AI endpoints. Storing keys in Vault prevents hardcoding them in application code or configuration files, reducing the risk of exposure and enabling automated rotation and access control via IAM policies.

Exam trap

The trap here is that candidates often confuse logging (Option A) with a security control, or assume that IP restrictions (Option B) are sufficient for securing an AI endpoint, when in fact OCI emphasizes private endpoints and authentication as best practices.

245
MCQmedium

A healthcare organization plans to deploy a RAG application on OCI that handles sensitive patient data. They require that all LLM inference and embedding processing happen within a controlled environment to avoid data leakage to public endpoints. Which OCI feature should they use?

A.OCI Data Labeling
B.OCI Vault
C.OCI Data Masking
D.OCI Dedicated AI Cluster
AnswerD

Dedicated AI Cluster provides isolated compute for AI workloads.

Why this answer

OCI Dedicated AI Cluster provides a fully isolated, single-tenant infrastructure for running AI workloads, including LLM inference and embedding processing. This ensures all data processing occurs within a controlled environment without traversing public endpoints, meeting the healthcare organization's requirement to prevent data leakage of sensitive patient data.

Exam trap

The common trap in Oracle exams is that candidates may confuse data protection features like OCI Vault (encryption) or OCI Data Masking with the need for a private compute environment, failing to recognize that OCI Dedicated AI Cluster is the only option that physically isolates the AI processing pipeline from public networks.

How to eliminate wrong answers

Option A is wrong because OCI Data Labeling is a service for creating labeled datasets for machine learning, not for controlling where inference or embedding processing occurs. Option B is wrong because OCI Vault is a key management service for storing and managing encryption keys and secrets, not for isolating compute environments for AI workloads. Option C is wrong because OCI Data Masking is used to redact or obfuscate sensitive data in databases and files, not to provide a private compute environment for LLM inference.

246
MCQmedium

A data scientist is creating a fine-tuning job in OCI Generative AI. They have prepared a JSONL dataset with prompt/completion pairs. What is the correct format for each line in the JSONL file?

A.{"input": "What is OCI?", "output": "Oracle Cloud Infrastructure is a cloud computing platform."}
B.{"prompt": "What is OCI?", "completion": "Oracle Cloud Infrastructure is a cloud computing platform."}
C.{"text": "What is OCI?", "label": "Oracle Cloud Infrastructure is a cloud computing platform."}
D.{"question": "What is OCI?", "answer": "Oracle Cloud Infrastructure is a cloud computing platform."}
AnswerB

This is the correct JSON format with 'prompt' and 'completion' fields.

Why this answer

The OCI Generative AI fine-tuning service expects JSONL files where each line contains exactly the keys "prompt" and "completion". This format matches the service's internal training pipeline, which maps the prompt to the input and the completion to the expected output during supervised fine-tuning.

Exam trap

The trap here is that candidates often confuse the key names with those used in other AI services (like OpenAI's fine-tuning which uses "prompt" and "completion" as well, but OCI's documentation explicitly requires these exact keys, and the exam tests attention to the specific OCI schema).

How to eliminate wrong answers

Option A is wrong because it uses "input" and "output" keys, which are not recognized by the OCI Generative AI fine-tuning API; the service requires the exact key names "prompt" and "completion". Option C is wrong because it uses "text" and "label" keys, which are commonly used in classification tasks but not for prompt/completion fine-tuning in OCI Generative AI. Option D is wrong because it uses "question" and "answer" keys, which are not the expected schema; the service strictly enforces the "prompt" and "completion" key names for JSONL lines.

247
Multi-Selecthard

A prompt engineer is using the ReAct pattern to enable the model to reason and act (e.g., call tools). Which THREE components are essential in the prompt to implement ReAct correctly?

Select 3 answers
A.A stop sequence after each action
B.A system prompt that sets a high temperature
C.A list of available actions or tools with descriptions
D.A scratchpad where the model can write intermediate thoughts
E.A few-shot example showing the Thought/Action/Observation loop
AnswersC, D, E

The model must know which actions it can take.

Why this answer

ReAct requires a scratchpad to record reasoning, a set of available actions (tools), and a format showing how to interleave reasoning and actions.

248
MCQhard

A company needs to classify customer support tickets into 20 categories. They have a labeled dataset of 50,000 examples. They want to use OCI Generative AI Embedding API to generate embeddings, then train a classifier. Which input type should they use for the embedding API when processing the training examples?

A.search_query
B.clustering
C.search_document
D.classification
AnswerD

classification input type is designed to produce embeddings that improve classifier performance.

Why this answer

For training a classifier, the 'classification' input type is optimized to generate embeddings that work well for classification tasks. Other types are for different use cases.

249
MCQhard

During multi-turn conversation with an OCI GenAI model, the model repeats user messages from earlier turns. What is the most likely cause?

A.Low top-p
B.High temperature
C.Low presence penalty
D.High frequency penalty
AnswerC

Low presence penalty means the model is less penalized for repeating topics, leading to repetition.

Why this answer

A low presence penalty reduces the model's incentive to avoid repeating previously mentioned content. In multi-turn conversations, this can cause the model to echo user messages from earlier turns because the penalty is too weak to discourage repetition of tokens that have already appeared in the context window.

Exam trap

Oracle often tests the distinction between presence penalty (which penalizes any occurrence) and frequency penalty (which penalizes based on count), leading candidates to mistakenly think a high frequency penalty causes repetition when it actually prevents it.

How to eliminate wrong answers

Option A is wrong because low top-p limits the cumulative probability mass for token sampling, which reduces diversity but does not directly cause repetition of earlier user messages; it may instead make outputs more deterministic. Option B is wrong because high temperature increases randomness in token selection, which can lead to more creative or even nonsensical outputs, not specifically the repetition of prior user messages. Option D is wrong because a high frequency penalty actively discourages the model from using tokens that have already appeared, which would reduce repetition, not cause it.

250
Multi-Selectmedium

Which TWO actions are best practices when deploying a RAG application using OCI OpenSearch and OCI Generative AI?

Select 2 answers
A.Embed every document chunk in real-time during query processing.
B.Implement a reranker to improve the relevance of retrieved documents.
C.Use very small chunk sizes (e.g., 50 tokens) to maximize granularity.
D.Monitor query latency and adjust the number of retrieved documents accordingly.
E.Set the LLM temperature to 1.5 to encourage diverse outputs.
AnswersB, D

Improves precision.

Why this answer

Implementing a reranker improves retrieval precision by re-scoring the top-k documents from the initial vector search using a cross-encoder model, which captures deeper semantic relevance than cosine similarity alone. In OCI OpenSearch, this is typically done via a post-processing step with OCI Generative AI or a dedicated reranking model, ensuring only the most contextually relevant chunks are passed to the LLM for generation.

Exam trap

Oracle often tests the misconception that real-time embedding (Option A) is efficient for RAG, when in fact pre-computed embeddings are standard, and that very small chunks (Option C) improve granularity, whereas they actually harm context coherence and retrieval quality.

251
MCQeasy

A company wants to use OCI Generative AI service to generate marketing copy that adheres to brand guidelines. Which technique should they use?

A.Use model distillation
B.Use prompt engineering with a pre-trained model
C.Use knowledge distillation
D.Fine-tune the model with brand-specific data
AnswerD

Correct: Fine-tuning adjusts model weights to match brand style and guidelines.

Why this answer

Fine-tuning a pre-trained model with brand-specific data (Option D) is the correct approach because it adjusts the model's weights to align with the company's unique brand guidelines, tone, and vocabulary. This supervised learning process ensures the generated marketing copy consistently adheres to specific requirements, unlike prompt engineering which relies on ephemeral instructions that may not reliably enforce brand constraints.

Exam trap

Oracle often tests the distinction between prompt engineering (which is temporary and instruction-based) and fine-tuning (which permanently alters model behavior), leading candidates to choose prompt engineering because it seems simpler, but it fails to guarantee adherence to brand guidelines.

How to eliminate wrong answers

Option A is wrong because model distillation is a technique to compress a large model into a smaller, faster one, not to adapt outputs to brand guidelines. Option B is wrong because prompt engineering with a pre-trained model can guide outputs but does not permanently embed brand-specific rules; the model may still deviate from guidelines without fine-tuned weights. Option C is wrong because knowledge distillation transfers knowledge from a teacher model to a student model for efficiency, not for customizing outputs to brand-specific data.

252
Multi-Selectmedium

A prompt engineer is testing a new prompt for a Q&A system. The prompt includes ambiguous wording that causes the model to answer inconsistently. Which TWO steps should the engineer take to resolve this? (Choose two.)

Select 2 answers
A.Use A/B testing between the original and rewritten prompt to measure improvement
B.Increase the temperature to encourage diverse responses
C.Test the prompt with a diverse set of inputs to identify remaining ambiguities
D.Reduce the max tokens to force shorter answers
E.Rewrite the prompt to remove ambiguous terms and provide clear context
AnswersC, E

Testing with varied inputs helps reveal where the prompt is still unclear.

Why this answer

Clarifying the prompt removes ambiguity. Testing with diverse inputs helps uncover edge cases. A/B testing would compare variants but does not directly fix ambiguity.

Temperature reduction does not clarify instructions.

253
MCQhard

A machine learning engineer is deploying a fine-tuned Llama 2 model on OCI Data Science model deployment. The deployment fails with an error: 'Model artifact exceeds the maximum allowed size of 10 GB.' The model files total 12 GB. What is the best approach to resolve this?

A.Store the model in Object Storage and reference it in the deployment configuration
B.Use a different model that is smaller than 10 GB
C.Increase the model deployment artifact size limit via a service request
D.Compress the model artifact to under 10 GB using gzip
AnswerA

Object Storage allows large models and is supported by model deployment.

Why this answer

OCI Data Science model deployment has a hard limit of 10 GB for the model artifact uploaded directly. By storing the model in Object Storage and referencing it in the deployment configuration, you bypass this limit entirely, as the deployment service can load the model from Object Storage at runtime without requiring the artifact to be part of the deployment package.

Exam trap

Oracle often tests the misconception that you can increase service limits via a support ticket, but for model artifact size, the limit is architectural and not adjustable; candidates may also incorrectly assume compression solves the issue without considering decompression at runtime.

How to eliminate wrong answers

Option B is wrong because it suggests a workaround that may not be feasible; the engineer has already fine-tuned a specific Llama 2 model, and switching to a smaller model would require retraining and may not meet business requirements. Option C is wrong because the 10 GB artifact size limit is a hard platform constraint that cannot be increased via a service request; OCI does not allow raising this limit for model deployments. Option D is wrong because compressing the artifact with gzip does not reduce the actual size of the model files when decompressed; the deployment service would need to decompress them, and the uncompressed size would still exceed the 10 GB limit, causing the same error.

254
MCQhard

A company uses an LLM to generate product descriptions. The outputs are consistently too verbose and include irrelevant details. The prompt includes a simple instruction: 'Describe the product.' Which adjustment to the prompt is most likely to yield concise, relevant descriptions?

A.Set temperature to 0.
B.Increase max_tokens to 500.
C.Add constraints like 'Max 30 words. Focus on key features.'
D.Include a few examples of desired short descriptions.
AnswerC

Explicit constraints directly limit length and scope.

Why this answer

Adding explicit constraints like 'Max 30 words. Focus on key features.' directly instructs the LLM to limit verbosity and prioritize relevant details. This technique, known as prompt engineering with constraints, is the most effective way to control output length and content without altering model parameters or relying on examples that may not generalize.

Exam trap

Oracle often tests the misconception that adjusting model parameters (temperature or max_tokens) is the primary way to control output quality, when in fact prompt engineering with explicit constraints is a more direct and reliable method for achieving specific formatting or length requirements.

How to eliminate wrong answers

Option A is wrong because setting temperature to 0 makes the model deterministic (greedy decoding), which reduces randomness but does not inherently shorten or focus the output—it may still produce verbose descriptions. Option B is wrong because increasing max_tokens to 500 actually allows the model to generate longer responses, which is counterproductive to achieving concise descriptions. Option D is wrong because including a few examples (few-shot prompting) can guide the style but does not guarantee brevity; the model may still extrapolate irrelevant details or exceed the desired length without explicit constraints.

255
Multi-Selectmedium

A data scientist is building a text summarization system using an LLM. They want to evaluate the model's output against human-written summaries. Which TWO metrics are most appropriate for this evaluation? (Choose two.)

Select 2 answers
A.Human evaluation rubrics
B.BLEU
C.ROUGE
D.Perplexity
E.BERTScore
AnswersC, E

ROUGE is recall-oriented and widely used for summarization evaluation.

Why this answer

ROUGE is recall-oriented and measures overlap of n-grams, making it a standard metric for summarization. BERTScore measures semantic similarity using embeddings, which can capture meaning even when wording differs. BLEU is more for translation, perplexity measures fluency, and human evaluation is qualitative but not a metric.

256
MCQmedium

An e-commerce company fine-tuned a Cohere Command model on their product catalog to generate product descriptions. During inference, they notice the model outputs are too repetitive: it often repeats similar phrases across different products, and the descriptions lack diversity. The team wants to increase the variety of the generated text without sacrificing relevance. They are currently using temperature=0.8, top_p=0.9, frequency_penalty=0, and presence_penalty=0. Which parameter adjustment should they make to most effectively increase diversity?

A.Decrease temperature from 0.8 to 0.5.
B.Set frequency_penalty to a negative value (e.g., -0.5).
C.Increase max_tokens from 200 to 500.
D.Increase top_p from 0.9 to 0.95.
AnswerD

Higher top_p includes more tokens in the sampling pool, increasing diversity.

Why this answer

Increasing top_p from 0.9 to 0.95 expands the nucleus of tokens considered during sampling, allowing the model to select from a wider set of plausible next tokens. This directly increases output diversity while still maintaining relevance, as tokens outside the top 90% probability mass are now included. The current settings already have moderate temperature and no penalties, so broadening top_p is the most effective single adjustment to reduce repetitiveness.

Exam trap

Oracle often tests the misconception that increasing temperature always increases diversity, when in fact decreasing temperature reduces randomness, and the most effective lever for diversity in a fine-tuned model is often adjusting top-p or adding a positive frequency penalty.

How to eliminate wrong answers

Option A is wrong because decreasing temperature from 0.8 to 0.5 makes the model more deterministic, reducing randomness and likely increasing repetitiveness, which is the opposite of the desired outcome. Option B is wrong because setting frequency_penalty to a negative value (e.g., -0.5) encourages the model to repeat tokens, exacerbating the repetitiveness problem rather than solving it. Option C is wrong because increasing max_tokens from 200 to 500 only extends the length of generated text; it does not alter the sampling strategy, so the model will continue to repeat phrases within the longer output.

257
MCQmedium

A financial institution uses OCI GenAI to power a customer support chatbot. The compliance team requires that responses are strictly consistent with regulatory guidelines and approved responses. The company has a curated set of question-answer pairs that cover common scenarios. They want to ensure that the chatbot never deviates from these approved answers. The data science team is considering various approaches to enforce this consistency. Which approach is most effective?

A.Few-shot prompting with three example responses in every query.
B.Fine-tuning the model on the curated dataset of question-answer pairs.
C.Using a large context window to include all regulatory guidelines in the prompt.
D.Setting a low temperature (0.1) to make outputs deterministic.
AnswerB

Fine-tuning adapts the model to mimic the approved responses, providing strong consistency.

Why this answer

Fine-tuning the model on the curated dataset of approved responses teaches the model to output similar responses for related questions, ensuring consistency. Option A is wrong because few-shot prompting may fail for unseen variations and does not guarantee strict adherence. Option C is wrong because using a large context window does not enforce specific content.

Option D is wrong because setting a low temperature reduces randomness but does not guarantee the model will choose approved responses.

258
MCQmedium

A company is using Oracle AI Vector Search in Oracle Database 23ai for semantic search over product descriptions. They need to create an index that supports approximate nearest neighbor search with high recall and moderate indexing time. Which index type and parameters should they choose?

A.Exact nearest neighbor search without an index
B.HNSW index with default parameters
C.No index — rely on full table scan
D.IVF index with a large number of centroids
AnswerB

HNSW typically provides high recall and moderate indexing time compared to exact search, making it suitable for production semantic search.

Why this answer

HNSW (Hierarchical Navigable Small World) indexes offer high recall and faster search times, but building the index takes longer. IVF (Inverted File) indexes index faster but may have lower recall unless tuned. For high recall and moderate indexing time, HNSW is preferred because it provides better accuracy at the cost of longer build time.

259
MCQhard

A team fine-tuned a Cohere Command R model using the T-Few technique on a dataset of JSONL prompt/completion pairs. After deployment, they observe that the model's responses are too repetitive. Which parameter adjustment in the OCI Generative AI Playground would BEST address this issue?

A.Increase frequency penalty
B.Increase max tokens
C.Decrease temperature
D.Add a stop sequence
AnswerA

Increasing the frequency penalty reduces the likelihood of the model repeating the same tokens or phrases by subtracting a penalty value from the logits of tokens that have already appeared in the generated text. This directly addresses the repetitive responses observed after fine-tuning with the T-Few technique, which can overfit to common patterns in the JSONL prompt/completion pairs.

Why this answer

Increasing the frequency penalty discourages the model from repeating tokens that have already appeared, reducing repetitive outputs. Temperature and max tokens address randomness and length, not repetition specifically. Stop sequences are for terminating generation early.

260
MCQeasy

A data scientist wants to deploy a fine-tuned LLM on OCI for inference with low latency. Which OCI service should they use?

A.OCI Data Science Notebook Session
B.OCI Generative AI Service (Dedicated AI Cluster)
C.OCI Data Flow
D.OCI Functions
AnswerB

Dedicated AI Cluster is optimized for low-latency inference with reserved resources.

Why this answer

B is correct because OCI Generative AI Service with a Dedicated AI Cluster provides a managed, high-throughput, low-latency inference endpoint for fine-tuned LLMs. It leverages GPU-accelerated infrastructure and optimized serving stacks (e.g., vLLM, TensorRT-LLM) to minimize response times, making it ideal for production inference workloads.

Exam trap

The trap here is that candidates confuse development environments (Notebook Sessions) or general-purpose serverless compute (Functions) with purpose-built inference services, overlooking the need for GPU-accelerated, managed inference endpoints for low-latency LLM deployment.

How to eliminate wrong answers

Option A is wrong because OCI Data Science Notebook Session is an interactive development environment for prototyping and training, not a production-grade inference endpoint; it lacks auto-scaling, load balancing, and dedicated GPU serving for low-latency inference. Option C is wrong because OCI Data Flow is a serverless Apache Spark service designed for batch and stream data processing, not for real-time LLM inference. Option D is wrong because OCI Functions is a serverless compute service for short-lived, stateless functions (max 5-minute timeout) and does not support GPU acceleration or persistent model serving required for low-latency LLM inference.

261
MCQmedium

A data scientist receives an error when calling the embed_text API: "InvalidRequest: input too long". What is the most likely cause and solution?

A.The model specified is not supported for embeddings; use a different model.
B.The input text exceeds the maximum token limit for the model; truncate the input.
C.The API request rate exceeds the tenancy limit; reduce the request rate.
D.The API key is invalid or expired; regenerate the key.
AnswerB

Embedding models have a fixed maximum input length.

Why this answer

The error 'InvalidRequest: input too long' indicates that the input text exceeds the maximum token limit for the embedding model. OCI Generative AI embedding models, like all transformer-based models, have a fixed context window (e.g., 512 or 1024 tokens). The solution is to truncate the input to fit within that limit, as the API will reject overly long inputs.

Exam trap

Oracle OCI GenAI exams often test the distinction between different error types (input length vs. rate limits vs. authentication) to see if candidates can map specific error messages to their root causes.

How to eliminate wrong answers

Option A is wrong because the error message specifically mentions 'input too long', not an unsupported model; unsupported models would return a different error like 'Model not found' or 'Invalid model'. Option C is wrong because rate limit errors typically return '429 Too Many Requests' or 'RateLimitExceeded', not 'InvalidRequest: input too long'. Option D is wrong because invalid or expired API keys return '401 Unauthorized' or 'Invalid API Key', not an input length error.

262
Multi-Selectmedium

An organization maintains a library of prompt templates for various use cases. Which three practices are essential for effective prompt management? (Choose three.)

Select 3 answers
A.Adjust temperature and top-p for each prompt variant
B.Store prompt templates in a shared, accessible repository
C.Define evaluation criteria to measure prompt performance
D.Conduct A/B testing on every prompt change
E.Version each prompt template and track changes
AnswersB, C, E

Central storage promotes reuse and consistency.

Why this answer

Versioning, storing templates in a central library, and establishing evaluation criteria are key to managing prompts. A/B testing is good but not a management practice; parameter tuning is separate.

263
MCQmedium

An organization needs to deploy a model that can both understand and generate text, such as for a translation task where the input is in English and output is in French. Which model architecture is most suitable?

A.Encoder-decoder (e.g., T5)
B.Decoder-only (e.g., GPT)
C.Encoder-only (e.g., BERT)
D.Mixture of Experts (MoE)
AnswerA

T5 is an encoder-decoder model specifically designed for text-to-text tasks like translation, where the full input is encoded and the decoder generates the output.

Why this answer

Encoder-decoder architectures like T5 are designed for sequence-to-sequence tasks. The encoder processes the input sequence, and the decoder generates the output sequence, making it ideal for translation.

264
Multi-Selecthard

Which THREE factors should be considered when choosing between fine-tuning a model and using a pre-trained model with prompt engineering? (Select three.)

Select 3 answers
A.Required response time
B.Size of available dataset
C.Internet connectivity
D.Available budget for compute resources
E.Need for domain-specific terminology
AnswersB, D, E

Fine-tuning requires a sufficiently large dataset; prompt engineering can work with few examples.

Why this answer

The size of the available dataset is a critical factor: fine-tuning requires a sufficiently large, labeled dataset (typically thousands of examples) to adjust model weights effectively, while prompt engineering can work with zero or few examples. If the dataset is too small, fine-tuning risks overfitting and poor generalization, making prompt engineering the safer choice.

Exam trap

Oracle often tests the misconception that response time or internet connectivity are decisive factors, when in reality the core trade-off is between data availability and the need for deep domain adaptation versus lightweight, zero-shot customization.

265
MCQhard

A healthcare startup is building a chatbot that retrieves patient treatment guidelines using OCI Generative AI Service and OCI OpenSearch. They require that all retrieved documents are from approved sources only and that the system can explain which source was used for each response. Which combination of features should they implement?

A.Add a metadata filter for source_type='approved' in the retrieval step and include document IDs in the context for the model.
B.Rely on the vector search's cosine similarity to rank approved sources higher.
C.Use prompt engineering to ask the model to ignore non-approved sources.
D.Reduce the top-K value to limit the number of retrieved documents.
AnswerA

Metadata filtering enforces source restriction; document IDs provide provenance.

Why this answer

It directly addresses both requirements: a metadata filter on `source_type='approved'` ensures only approved documents are retrieved from OpenSearch, and including document IDs in the context allows the model to cite the specific source for each response. This approach enforces access control at the retrieval layer while providing traceability, which is essential for compliance in healthcare applications.

Exam trap

The trap here is that candidates may assume semantic similarity or prompt engineering alone can enforce access control, but in RAG systems, retrieval-layer filtering is the only reliable way to restrict document access before the model sees the content.

How to eliminate wrong answers

Option B is wrong because cosine similarity measures semantic relevance, not source approval status; approved and non-approved documents can be equally similar to a query, so ranking by similarity alone cannot guarantee that only approved sources are used. Option C is wrong because prompt engineering cannot reliably filter out non-approved sources; the model may still see and inadvertently use non-approved content in its context, and it has no inherent mechanism to verify source approval. Option D is wrong because reducing the top-K value limits the number of retrieved documents but does not enforce any approval criterion; non-approved documents can still appear in the top-K results if they are semantically similar.

266
Multi-Selecthard

An enterprise is deploying a generative AI model that must comply with data residency regulations. Which two configurations should they implement? (Select TWO.)

Select 2 answers
A.Set up OCI IAM policies to prevent data egress from the region for the model's resources
B.Enable OCI Logging for all API calls
C.Use OCI Object Storage with cross-region replication for redundancy
D.Store encryption keys in an OCI Vault in a different region
E.Deploy the dedicated AI cluster in the region that meets data residency requirements
AnswersA, E

Correct: IAM policies can restrict access to resources from outside the region.

Why this answer

OCI IAM policies can explicitly deny data egress from a specific region, ensuring that the generative AI model's resources (such as training data, model artifacts, and inference endpoints) remain within the region that satisfies data residency regulations. This is achieved by writing policy statements that restrict the movement of data across regional boundaries, which is a direct control for compliance.

Exam trap

The trap here is that candidates often confuse data residency enforcement with monitoring or key management, mistakenly selecting logging (Option B) or cross-region replication (Option C) as compliance controls, when only IAM policies and regional deployment directly prevent data movement.

267
MCQhard

A prompt engineer is designing a ReAct pattern prompt to enable an LLM to use external tools. Which of the following is a key characteristic of the ReAct pattern?

A.The model uses a single tool call to answer the question
B.The model produces a chain of thought followed by an action, then observes the result and continues reasoning
C.The model generates a final answer without any intermediate steps
D.The model is fine-tuned specifically for tool use
AnswerB

Correct: ReAct alternates between reasoning, action, and observation.

Why this answer

ReAct interleaves reasoning steps (e.g., 'I need to find the current weather') with actions (e.g., 'Action: call weather API') and observations, allowing the model to reason and act in a cycle.

268
Multi-Selectmedium

A developer is building a LangChain RAG pipeline with OCI Generative AI. Which TWO components are needed to create embeddings from documents and store them for retrieval?

Select 2 answers
A.DocumentLoader
B.OCIGenAIEmbeddings
C.ChatOCIGenAI
D.Vector store (e.g., FAISS, OracleVS)
E.TextSplitter
AnswersB, D

This wraps OCI's embedding models.

Why this answer

OCIGenAIEmbeddings converts text into embeddings, and a vector store (e.g., FAISS, Chroma, OracleVS) stores those embeddings for similarity search. Document loaders and text splitters are used before embedding but are not part of the embedding/storage step itself.

269
MCQeasy

A company wants to build a customer support chatbot using OCI Generative AI. They have a large number of historical support tickets. Which approach is most effective for leveraging this data to improve the chatbot's responses?

A.Use a pre-loaded prompt template from the OCI console.
B.Fine-tune the Cohere Command model on the historical tickets using OCI Data Science.
C.Increase the temperature parameter to 1.0 to encourage diverse responses.
D.Use zero-shot prompting with the base model and include few-shot examples in the prompt.
AnswerB

Fine-tuning on the company's own support tickets adapts the model to the specific language, context, and resolutions, significantly improving response quality.

Why this answer

Fine-tuning the Cohere Command model on the historical support tickets using OCI Data Science is the most effective approach because it adapts the model's weights to the specific domain language, terminology, and resolution patterns found in the company's data. This supervised learning process creates a specialized model that can generate accurate, context-aware responses for customer support queries, unlike generic prompting methods that lack deep domain adaptation.

Exam trap

Oracle often tests the misconception that increasing temperature or using few-shot examples can substitute for fine-tuning when adapting a model to proprietary domain data, but in reality only fine-tuning modifies model weights to deeply learn domain-specific patterns from large datasets.

How to eliminate wrong answers

Option A is wrong because pre-loaded prompt templates in the OCI console are generic and not trained on the company's specific historical ticket data, so they cannot capture domain-specific nuances or improve response accuracy beyond basic instruction following. Option C is wrong because increasing the temperature parameter to 1.0 maximizes randomness in token selection, which reduces coherence and factual reliability—exactly the opposite of what is needed for a customer support chatbot that requires consistent, accurate answers. Option D is wrong because zero-shot prompting with few-shot examples only provides a few static examples in the context window, which does not modify the model's underlying weights and cannot match the depth of learning achieved by fine-tuning on thousands of historical tickets.

270
MCQmedium

Which of the following metrics is most suitable for evaluating a translation model's output against multiple reference translations?

A.ROUGE
B.Perplexity
C.BERTScore
D.BLEU
AnswerD

BLEU is the standard metric for machine translation.

Why this answer

BLEU (Bilingual Evaluation Understudy) is the most suitable metric for evaluating a translation model's output against multiple reference translations because it measures n-gram precision between the candidate translation and one or more reference translations. It directly quantifies how many words and phrases in the candidate match those in the references, making it the standard metric for machine translation tasks.

Exam trap

The trap here is that candidates often confuse ROUGE (recall-based) with BLEU (precision-based) or assume BERTScore's semantic matching is better for translation, but BLEU is explicitly the standard for multi-reference translation evaluation in the NLP community.

How to eliminate wrong answers

Option A is wrong because ROUGE (Recall-Oriented Understudy for Gisting Evaluation) focuses on recall of n-grams and is primarily designed for summarization evaluation, not translation. Option B is wrong because Perplexity measures how well a language model predicts a sequence of tokens, but it does not compare against reference translations and is not a direct evaluation metric for translation quality. Option C is wrong because BERTScore uses contextual embeddings from BERT to compute similarity between candidate and reference, but it is a semantic similarity metric and not specifically optimized for evaluating translation output against multiple references; BLEU remains the standard for this task.

271
MCQmedium

A company is deploying a fine-tuned Cohere model on OCI Generative AI service for real-time inference. They need to ensure low latency even during demand spikes. Which configuration should they prioritize?

A.Enable model caching on the endpoint.
B.Use a dedicated AI cluster for the endpoint.
C.Use streaming responses.
D.Increase the max tokens parameter.
AnswerB

A dedicated AI cluster with autoscaling ensures consistent low latency under variable load.

Why this answer

A dedicated AI cluster provides isolated compute resources (GPUs) that are not shared with other tenants or workloads, ensuring consistent low latency even under demand spikes. This is critical for real-time inference because shared endpoints can experience resource contention and throttling during high traffic, while a dedicated cluster guarantees predictable performance.

Exam trap

Oracle often tests the misconception that caching or streaming alone can solve latency under load, when in fact only dedicated compute resources guarantee isolation and consistent performance during demand spikes.

How to eliminate wrong answers

Option A is wrong because model caching reduces latency for repeated requests by storing intermediate results, but it does not prevent resource contention during demand spikes; it only helps with cache hits, not with ensuring low latency under sustained high load. Option C is wrong because streaming responses improve perceived latency by sending tokens as they are generated, but they do not address the underlying compute resource availability or prevent queuing delays during spikes. Option D is wrong because increasing the max tokens parameter increases the maximum output length, which can actually increase latency per request and does nothing to handle demand spikes or resource contention.

272
Multi-Selecthard

A company is deploying a generative AI model on OCI for an internal application that must comply with strict security policies. The model will be accessed by a limited group of users. Which three actions should the administrator take to ensure security? (Choose three.)

Select 3 answers
A.Expose the model endpoint to the internet for ease of access
B.Deploy the model in a private VCN subnet
C.Use IAM policies to restrict model endpoint access to specific users
D.Disable audit logging to minimize storage costs
E.Store model authentication keys in OCI Vault
AnswersB, C, E

A private subnet ensures the endpoint is not publicly accessible.

Why this answer

Deploying the model in a private VCN subnet ensures that the model endpoint is not exposed to the internet, which is a fundamental security requirement for compliance with strict security policies. By placing the model in a private subnet, all traffic must traverse through a bastion host, VPN, or FastConnect, providing network isolation and reducing the attack surface. This aligns with OCI's shared responsibility model where the customer controls network security.

Exam trap

The trap here is that candidates may think exposing the endpoint to the internet is acceptable if IAM policies are used, but network isolation (private subnet) is a separate and mandatory layer of defense that cannot be replaced by IAM alone.

273
MCQhard

An LLM application generates product descriptions. The output is sometimes repetitive (e.g., 'innovative' appears multiple times). Which parameter adjustment is MOST likely to reduce this repetition without harming creativity?

A.Increase temperature from 0.7 to 1.0
B.Set frequency penalty to a positive value, e.g., 0.3
C.Decrease max tokens from 200 to 100
D.Increase top-k from 40 to 80
AnswerB

Frequency penalty directly penalizes tokens that have already been used, reducing repetition.

Why this answer

The frequency penalty reduces the likelihood of repeating tokens that have already appeared. A moderate penalty (e.g., 0.3) discourages repetition while allowing creative word choices. Temperature and top-k affect randomness, not repetition.

274
Multi-Selectmedium

A company is using OCI Generative AI Agents to build a customer support assistant. They have uploaded product manuals to OCI Object Storage. Which two components are required to create the agent? (Select TWO)

Select 2 answers
A.Create a Generative AI Agent using the knowledge base
B.Create an endpoint via InferenceClient
C.Create a knowledge base from the Object Storage bucket
D.Upload data directly to the agent (not via knowledge base)
E.Provision a Dedicated AI Cluster
AnswersA, C

The agent uses the knowledge base to answer questions.

Why this answer

A Generative AI Agent requires a knowledge base to provide domain-specific context for answering queries. The knowledge base is the source of truth that the agent uses to ground its responses, and it must be explicitly created and associated with the agent. Without a knowledge base, the agent would lack the product manual data needed for customer support.

Exam trap

A common misconception in OCI is that you can directly upload data to a Generative AI Agent or that you must manually create an endpoint, but in reality the agent relies on a knowledge base for data ingestion and uses a managed endpoint automatically.

275
MCQhard

A data scientist observes that their fine-tuned LLM performs well on training data but generates repetitive and dull responses in production. What is the most likely cause and best solution?

A.The model is overfitted; apply stronger regularization
B.The temperature is set too low; increase temperature during inference
C.The training data lacks diversity; add more varied examples
D.The model has too many layers; reduce model size
AnswerB

Low temperature makes outputs deterministic and repetitive; increasing it adds variability.

Why this answer

The model's repetitive and dull responses indicate that the temperature parameter is too low, causing the model to always select the most probable tokens, leading to deterministic and monotonous outputs. Increasing temperature during inference introduces randomness into token sampling, allowing for more diverse and creative responses. This is a common issue in production LLMs where low temperature settings optimized for training metrics fail to produce engaging real-world outputs.

Exam trap

Oracle often tests the misconception that poor production performance is always due to overfitting or data issues, when in fact inference-time hyperparameters like temperature are the direct cause of repetitive/dull outputs.

How to eliminate wrong answers

Option A is wrong because overfitting would cause poor generalization to new inputs, not specifically repetitive/dull outputs; regularization reduces overfitting but does not address the deterministic token selection caused by low temperature. Option C is wrong because while training data diversity affects model knowledge, the described symptom of repetitive outputs in production despite good training performance points to inference-time sampling issues, not data diversity. Option D is wrong because having too many layers might cause overfitting or computational inefficiency, but it does not directly cause repetitive or dull responses; reducing model size would not fix the temperature-related sampling behavior.

276
MCQhard

A security administrator wrote the above IAM policy for a compartment named MyCompartment. Users in the GenerativeAIUsers group can successfully list dedicated AI clusters and models in MyCompartment, but when they try to create an inference endpoint using a model from a different compartment (SharedModels), they get an authorization error. What is the most likely missing policy statement?

A.ALLOW GROUP GenerativeAIUsers TO MANAGE generative-ai-models IN COMPARTMENT SharedModels
B.ALLOW GROUP GenerativeAIUsers TO USE generative-ai-family IN TENANCY
C.ALLOW GROUP GenerativeAIUsers TO USE generative-ai-dedicated-ai-clusters IN COMPARTMENT SharedModels
D.ALLOW GROUP GenerativeAIUsers TO USE generative-ai-models IN COMPARTMENT SharedModels
AnswerD

This allows them to use models from SharedModels compartment.

Why this answer

The error occurs because the user has permission to list models in MyCompartment but not to use a model from SharedModels when creating an inference endpoint. The missing policy must grant the USE permission on generative-ai-models in the SharedModels compartment, as creating an endpoint requires the ability to reference and use the model resource from that compartment. Option D correctly provides this permission.

Exam trap

Oracle often tests the distinction between 'read' and 'use' permissions, where candidates mistakenly think listing models (read) is sufficient to use them in another resource creation, but OCI requires the 'use' verb for referencing a resource across compartments.

How to eliminate wrong answers

Option A is wrong because it grants MANAGE permission, which is excessive; the user only needs USE permission to reference the model for creating an endpoint. Option B is wrong because it grants USE on the entire generative-ai-family at the tenancy level, which is too broad and not scoped to the specific model resource needed from SharedModels. Option C is wrong because it grants USE on generative-ai-dedicated-ai-clusters in SharedModels, but the error is about using a model, not a dedicated AI cluster.

277
MCQmedium

A data scientist is designing a prompt to generate a structured report with sections for Summary, Findings, and Recommendations. Which output format specification in the prompt would be MOST effective?

A."Write a report with three sections: Summary, Findings, Recommendations."
B."Provide the output in JSON format with keys: 'summary', 'findings', and 'recommendations'."
C."Return the report in bullet points."
D."Output the report as a markdown document with headings."
AnswerB

Explicit JSON key specification yields a structured, easily parsed output.

Why this answer

Specifying JSON output with clear keys ensures the model returns a structured, machine-parseable result. Natural language descriptions are ambiguous, and markdown may not be reliably parsed.

278
Multi-Selectmedium

A data scientist is debugging a RAG system where the generated answers are not relevant to the retrieved documents. Which TWO factors are MOST likely causing this issue?

Select 2 answers
A.The retriever is returning irrelevant chunks due to poor embeddings or low similarity threshold
B.The generation model is not conditioned on the retrieved chunks, possibly because the prompt does not instruct it to use them
C.The context window of the generation model is smaller than the retrieved chunks
D.The temperature is set too low, making outputs deterministic
E.The chunking strategy produces chunks that are too small, losing context
AnswersA, B

Irrelevant chunks lead to irrelevant answers; this is a common cause.

Why this answer

If the retriever returns irrelevant chunks or the generation model ignores the context, the answers will be off-topic. Checking retrieval relevance and prompt instruction is key.

279
MCQhard

A global enterprise is deploying a generative AI application that requires high availability across multiple OCI regions. The application must automatically fail over to a secondary region if the primary region becomes unavailable. What is the recommended architecture to achieve this?

A.Deploy endpoints in two regions behind an OCI Load Balancer with cross-region failover
B.Deploy OCI Generative AI endpoints in two regions and use a global DNS round-robin
C.Use OCI Streaming to replicate requests between regions
D.Use DNS failover with a single endpoint in the primary region
AnswerA

OCI Load Balancer can route traffic to a backup region when primary is unhealthy.

Why this answer

OCI Load Balancer supports cross-region failover by distributing traffic across backend sets in multiple regions, enabling automatic failover to a secondary region when the primary region becomes unavailable. This architecture ensures high availability for generative AI applications by leveraging health checks and failover policies at the load balancer level, which is the recommended approach for multi-region active-passive setups.

Exam trap

Oracle often tests the misconception that DNS-based solutions (like round-robin or simple failover) provide automatic failover with health checks, but in OCI, DNS failover requires manual intervention or additional services like Traffic Management Steering, whereas OCI Load Balancer natively supports automatic cross-region failover.

How to eliminate wrong answers

Option B is wrong because global DNS round-robin does not provide automatic failover; it distributes traffic statically and cannot detect regional outages, leading to continued traffic to an unavailable endpoint. Option C is wrong because OCI Streaming is a messaging service for real-time data ingestion and replication, not a traffic routing or failover mechanism for application endpoints. Option D is wrong because DNS failover with a single endpoint in the primary region lacks a secondary region for failover, offering no high availability if the primary region fails.

280
Multi-Selecteasy

A data scientist is preparing to fine-tune a foundation model on OCI. Which two actions should they take to optimize costs? (Select TWO.)

Select 2 answers
A.Use the smallest model that meets accuracy requirements
B.Use a single OCPU shape to minimize per-hour cost
C.Use spot preemptible instances to save on compute
D.Monitor fine-tuning progress and stop early if validation loss plateaus
E.Store training data in Archive Storage to reduce storage costs
AnswersA, D

Correct: Smaller models require less compute and memory.

Why this answer

Using the smallest model that meets accuracy requirements directly reduces the number of parameters and computational operations required during fine-tuning. On OCI, larger models consume significantly more GPU memory and compute hours, so selecting the minimal viable model minimizes both training time and associated costs. This aligns with cost optimization best practices for generative AI workloads.

Exam trap

Oracle often tests the misconception that spot/preemptible instances are universally cost-effective for all AI workloads, but in OCI, they are not supported for interactive or stateful fine-tuning jobs, making Option C a classic distractor.

281
MCQhard

During fine-tuning a model using T-Few in OCI Generative AI, the job fails with a 'dataset format error'. The training dataset is a JSONL file. Which of the following is the MOST likely cause?

A.The file contains more than 1000 lines
B.Some lines have an extra 'metadata' key
C.The completion field contains more than 1024 tokens
D.The file uses Unix line endings
AnswerB

The dataset format expects only 'prompt' and 'completion'. Extra keys like 'metadata' are not allowed and cause validation errors.

Why this answer

Each JSONL line must contain exactly 'prompt' and 'completion' keys. Extra keys or missing keys cause format errors.

282
MCQmedium

Which evaluation metric is designed to measure the overlap of n-grams between a generated summary and a reference summary, focusing on recall of content words?

A.Perplexity
B.ROUGE
C.BERTScore
D.BLEU
AnswerB

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures recall of n-grams between generated and reference summaries.

Why this answer

ROUGE-N measures n-gram recall (how many n-grams from the reference appear in the generated text). BLEU measures n-gram precision. BERTScore leverages contextual embeddings.

Perplexity measures likelihood under the model.

283
MCQeasy

What is the PRIMARY purpose of the 'stop sequences' parameter in text generation?

A.To prevent the model from generating offensive content
B.To specify tokens that the model should avoid using
C.To encourage the model to include specific phrases in the output
D.To define sequences that, when generated, cause the model to cease generation
AnswerD

Stop sequences like ' ' or '###' halt generation at that point.

Why this answer

Stop sequences tell the model when to stop generating further tokens, allowing control over output length or termination upon specific strings.

284
MCQmedium

A developer uses the ReAct agent in LangChain with a calculator tool and a search tool. The agent receives the question: 'What is the population of Paris multiplied by 3?' The agent first calls the search tool to find the population, then calls the calculator tool to multiply it by 3. Which component is responsible for deciding the sequence of tool calls?

A.The LLM
B.The Memory
C.The Tool
D.The AgentExecutor
AnswerD

The AgentExecutor repeatedly invokes the LLM, executes tool calls, and passes results back until the agent stops.

Why this answer

The AgentExecutor orchestrates the agent's reasoning loop: it calls the LLM which decides the next action (tool call) based on the prompt (ReAct pattern). The LLM generates actions, the AgentExecutor executes them and feeds results back to the LLM until a final answer is produced. Tools are just functions, the Agent defines the prompt, and Memory stores history.

285
MCQmedium

A team is implementing a conversational chatbot that needs to remember a user's previous messages within the same session. They are using LangChain with OCI Generative AI. Which memory type and persistence approach should they choose for session-only memory?

A.ConversationBufferMemory persisted in a vector store
B.ConversationSummaryMemory stored in a local file
C.ConversationBufferMemory stored in-memory
D.ConversationSummaryMemory persisted in Oracle Database
AnswerC

Buffer memory retains the full conversation history in Python memory, ideal for session-scoped chatbots that do not require long-term persistence.

Why this answer

ConversationBufferMemory stores the full message history in memory, suitable for short sessions. For session-only memory, in-memory persistence is sufficient; no external database is needed. Summary memory is more suited for long conversations where token limits are a concern.

286
MCQmedium

A developer is using LangChain to build a RAG pipeline with Oracle Database 23ai as the vector store. Which LangChain wrapper should they use to create embeddings and store them in the database?

A.OCIGenAI for embeddings and FAISS for storage
B.HuggingFaceEmbeddings for embeddings and OracleVS for storage
C.ChatOCIGenAI for embeddings and Chroma for storage
D.OCIGenAIEmbeddings for embeddings and OracleVS for storage
AnswerD

OCIGenAIEmbeddings generates embeddings using Oracle's embedding models, and OracleVS stores them in Oracle Database 23ai with native vector support.

Why this answer

OCIGenAIEmbeddings is the LangChain wrapper for Oracle's embedding models, and OracleVS connects to Oracle Database 23ai for vector storage. The other options either refer to incorrect wrappers or are not designed for this purpose.

287
MCQmedium

A team wants to use an LLM to answer questions about a private codebase that is updated hourly. They cannot afford to fine-tune every hour. Which OCI feature or approach is most suitable?

A.Use a long-context model with full codebase in prompt
B.Implement Retrieval-Augmented Generation (RAG) with a vector database
C.Fine-tune a model on the codebase daily
D.Use a smaller model with faster inference
AnswerB

RAG provides up-to-date retrieval without retraining.

Why this answer

Retrieval-Augmented Generation (RAG) with a vector database is the most suitable approach because it allows the LLM to answer questions about a frequently updated private codebase without retraining. RAG retrieves relevant code snippets from a vector index at query time, ensuring the model always has access to the latest code without the cost and latency of hourly fine-tuning.

Exam trap

The trap here is that candidates often assume fine-tuning is the only way to incorporate private or dynamic data, overlooking RAG's ability to provide real-time, cost-effective access to frequently updated information without retraining.

How to eliminate wrong answers

Option A is wrong because a long-context model with the full codebase in the prompt is impractical: the codebase is updated hourly and likely exceeds the model's context window, leading to truncation, high token costs, and degraded performance. Option C is wrong because fine-tuning a model daily (or hourly) is too expensive and time-consuming for a rapidly changing codebase, and it does not support real-time updates without retraining. Option D is wrong because using a smaller model with faster inference does not solve the core problem of accessing up-to-date private code; it only addresses inference speed, not knowledge freshness or retrieval.

288
MCQmedium

A practitioner is using a Cohere Command model on OCI for a translation task. They notice that the output is often incomplete and cuts off mid-sentence. Which parameter should they adjust to address this?

A.Temperature
B.Max tokens
C.Frequency penalty
D.Top-p
AnswerB

Max tokens sets the maximum length of the generated output.

Why this answer

The 'Max tokens' parameter controls the maximum length of the generated output. When a model cuts off mid-sentence, it means the token limit has been reached before the model could complete its response. Increasing this value allows the model to generate more tokens, thus completing the translation.

Exam trap

The 1Z0-1127 exam often tests the misconception that temperature or top-p controls output length, when in fact they only affect token selection probability and diversity, not the maximum number of tokens generated.

How to eliminate wrong answers

Option A is wrong because Temperature controls the randomness of the output, not the length; lowering it makes the model more deterministic but does not prevent truncation. Option C is wrong because Frequency penalty reduces repetition by penalizing tokens that have already appeared, but it does not affect the total number of tokens generated. Option D is wrong because Top-p (nucleus sampling) controls the cumulative probability threshold for token selection, influencing diversity, not the output length.

289
Multi-Selectmedium

Which TWO components are essential in a Retrieval-Augmented Generation (RAG) pipeline?

Select 2 answers
A.Chunking the documents into smaller pieces
B.Embedding the chunks into a vector space
C.Fine-tuning the LLM on the documents
D.Knowledge distillation
E.Beam search decoding
AnswersA, B

Chunking is necessary for indexing.

Why this answer

Chunking splits documents, embedding converts chunks to vectors, retrieval fetches relevant chunks, and generation produces the answer.

290
MCQhard

A team is fine-tuning a large language model for a domain-specific Q&A application. After fine-tuning, they observe that the model performs well on the training distribution but struggles with out-of-distribution (OOD) questions. Which approach would best improve OOD robustness?

A.Include a diverse set of examples from related domains in the fine-tuning dataset.
B.Use early stopping based on training loss to avoid overfitting.
C.Reduce the model size to prevent overfitting to the training data.
D.Increase the learning rate during fine-tuning to adapt faster to new patterns.
AnswerA

Diverse data improves generalization and OOD performance.

Why this answer

Including diverse examples from related domains (Option A) exposes the model to varied patterns, reducing overfitting to the training distribution and improving generalization to out-of-distribution inputs. Option B (early stopping) helps prevent overfitting on the validation set but does not specifically address OOD robustness. Option C (reducing model size) limits capacity, which can harm performance on both in-distribution and OOD data.

Option D (increasing learning rate) risks catastrophic forgetting and training instability, not OOD robustness.

291
Multi-Selecthard

A team is troubleshooting a chatbot that sometimes outputs harmful content despite having a system prompt with safety instructions. Which THREE measures should they implement to reduce the risk?

Select 3 answers
A.Reinforce the system prompt with explicit safety constraints and periodic reminders
B.Increase temperature to make output more random
C.Add more few-shot examples of safe responses
D.Use output filtering to scan and block harmful responses before showing to users
E.Implement input filtering to detect and block malicious prompt injection attempts
AnswersA, D, E

Strengthening the system prompt with detailed safety rules reduces the chance of harmful outputs.

Why this answer

Input filtering blocks malicious inputs, output filtering catches harmful responses before delivery, and system prompt reinforcement strengthens initial instructions. Adding more few-shot examples does not directly address safety.

292
Multi-Selecthard

Which THREE factors are important when designing a multi-turn conversational agent using OCI Generative AI Agents?

Select 3 answers
A.Always generate the longest possible response to be thorough.
B.Manage the context window size to avoid truncating important earlier messages.
C.Implement guardrails to detect and filter sensitive topics or harmful intents.
D.Disable logging to reduce latency and cost.
E.Enable session management to maintain conversation history across turns.
AnswersB, C, E

If the context window is too small, the agent may lose track of earlier parts of the conversation.

Why this answer

Managing the context window size is critical because OCI Generative AI Agents have a fixed token limit for the conversation history. If the context window is exceeded, the agent truncates the oldest messages, which can remove essential context from earlier turns, leading to incoherent or incorrect responses. Proper management ensures that the most relevant history is retained without exceeding the model's maximum input length.

Exam trap

Oracle often tests the misconception that longer responses are better for thoroughness, when in fact they degrade performance and user experience, and that disabling logging is a harmless optimization, whereas it removes critical observability and debugging capabilities.

293
MCQeasy

A developer is testing the OCI Generative AI API by sending a request to generate text using the Cohere Command R model. The request returns the following error: 'The model 'cohere.command-r-08-2024' is not available in this region. Please check the model availability in your region.' The developer is using the us-ashburn-1 region. What is the most likely cause of this error?

A.The request body format is incorrect.
B.The model is not deployed in the us-ashburn-1 region.
C.The model name is misspelled (e.g., 'cohere.command-r-08-2024' vs 'cohere.command-r-08-2024').
D.The API key used in the request is invalid.
AnswerB

Cohere Command R may not be available in all regions; check supported regions in OCI documentation.

Why this answer

The error message explicitly states that the model 'cohere.command-r-08-2024' is not available in the region. OCI Generative AI models are deployed regionally, and the Cohere Command R model is not available in the us-ashburn-1 (Ashburn) region. The developer must select a supported region, such as us-chicago-1, where this model is deployed.

Exam trap

Oracle often tests the misconception that model names must be perfectly spelled or that API keys are the cause of all errors, but here the trap is that candidates overlook regional availability and assume the error is due to a typo or authentication failure.

How to eliminate wrong answers

Option A is wrong because an incorrect request body format would typically result in a 400 Bad Request or validation error, not a model availability error. Option C is wrong because the model name in the error matches the one sent, so a misspelling would cause a different error (e.g., 'model not found'), not a region availability error. Option D is wrong because an invalid API key would result in a 401 Unauthorized or 403 Forbidden error, not a model availability error.

294
MCQhard

A practitioner needs to choose a pre-trained model for a sentiment analysis task on customer reviews. The model must be efficient for inference and capable of handling multiple languages. Which architecture is MOST suitable?

A.Encoder-only BERT model
B.Encoder-decoder T5 model
C.Decoder-only GPT model
D.Mixture of Experts model
AnswerD

Mixture of Experts models achieve efficiency through sparse activation and can specialize experts per language, making them ideal for multilingual sentiment analysis with fast inference.

Why this answer

Mixture of Experts (MoE) models are designed for efficient inference by activating only a subset of parameters per input, reducing computational cost. They can also support multiple languages by allocating different experts to different language patterns, making them highly suitable for multilingual sentiment analysis. In contrast, encoder-only BERT models are efficient but may not scale as well for multilingual tasks without large capacity, encoder-decoder models are optimized for sequence-to-sequence tasks, and decoder-only models are primarily generative and less efficient for classification.

Exam trap

Candidates may default to BERT as the standard for classification, but MoE offers better efficiency and multilingual support in modern architectures.

295
MCQmedium

A data scientist deployed a fine-tuned Llama 2 7B model on OCI Model Deployment with a single VM.GPU.A10.1 shape. Users report average latency of 3 seconds per request, which is too high for the intended real-time application. The model is used for short text generation (max 128 tokens). The data scientist wants to reduce per-request latency without significant accuracy loss. Which action would be most effective?

A.Increase the number of workers per replica
B.Increase the max_tokens parameter for the model
C.Enable response streaming for the model endpoint
D.Apply 4-bit quantization using AWQ
AnswerD

Quantization reduces model size and inference time with minimal accuracy loss.

Why this answer

4-bit quantization using AWQ reduces the model's memory footprint and computational requirements by compressing weights to 4-bit integers, which directly decreases inference latency on the VM.GPU.A10.1 shape. This technique preserves most of the model's accuracy while enabling faster token generation, making it the most effective single action for reducing per-request latency in a real-time short text generation scenario.

Exam trap

The trap here is that candidates confuse throughput improvements (Option A) or perceived latency (Option C) with actual per-request latency reduction, or mistakenly think increasing max_tokens (Option B) would help, when in fact it worsens the problem.

How to eliminate wrong answers

Option A is wrong because increasing the number of workers per replica does not reduce per-request latency; it only increases throughput by handling more concurrent requests, but each individual request still experiences the same inference time. Option B is wrong because increasing the max_tokens parameter would actually increase latency, as the model would generate more tokens per request, making the problem worse. Option C is wrong because enabling response streaming does not reduce the total time to generate the full response; it only sends tokens incrementally to the client, improving perceived latency but not actual end-to-end latency.

296
MCQeasy

Which of the following is a distinguishing feature of in-context learning compared to fine-tuning?

A.In-context learning modifies the model's weights based on examples
B.In-context learning does not update the model's weights; instead, examples are provided in the prompt
C.In-context learning is only possible with encoder-only models
D.In-context learning requires additional training on a labeled dataset
AnswerB

In-context learning uses examples in the prompt at inference time without any weight updates.

Why this answer

In-context learning does not update model weights; it provides examples in the prompt at inference time. Fine-tuning updates the model weights through additional training on a dataset.

297
MCQhard

You are a cloud architect at a global e-commerce company. The company is building a RAG-based product support chatbot using OCI Generative AI Service and OCI OpenSearch. The chatbot must answer customer questions in real-time by retrieving from a product knowledge base containing over 10 million documents. The current architecture uses a single vector index with all documents, and the LLM (Cohere Command R+) returns answers in English only. The team observes that queries from non-English customers often return irrelevant results, and the chatbot sometimes fails to generate answers within the 5-second SLA. The leadership wants to support 10 languages and reduce the average response time to under 3 seconds. You need to propose a solution that improves both relevance and latency. Which course of action should you take?

A.Increase the number of OCI OpenSearch nodes and upgrade the LLM to a faster variant.
B.Replace the embedding model with a multilingual model and partition the vector index by language to reduce search space.
C.Translate all non-English queries to English before retrieval and use an English-only embedding model.
D.Implement a caching layer for frequent queries and use a larger LLM for better accuracy.
AnswerB

Multilingual model improves relevance; partitioning improves latency.

Why this answer

Partitioning the vector index by language reduces the search space for each query, directly improving retrieval latency, while using a multilingual embedding model ensures that non-English queries are semantically matched to documents in their original language, improving relevance. This combination addresses both the 3-second SLA and the 10-language requirement without relying on translation, which introduces latency and potential loss of meaning.

Exam trap

The trap here is that candidates often assume translation is the simplest path to multilingual support, overlooking the latency and semantic drift it introduces, and fail to recognize that partitioning the index is a standard optimization for both relevance and speed in large-scale RAG systems.

How to eliminate wrong answers

Option A is wrong because simply scaling nodes and upgrading the LLM does not fix the root cause of irrelevant results for non-English queries—the embedding model remains English-only, so multilingual queries will still map poorly in vector space. Option C is wrong because translating all non-English queries to English before retrieval adds significant latency (often 200-500ms per translation) and can lose cultural or contextual nuances, making it unsuitable for a 3-second SLA and 10-language support. Option D is wrong because a caching layer only helps with repeated queries, not novel ones, and using a larger LLM increases inference latency, making it harder to meet the 3-second target; it also does not address the embedding mismatch for non-English content.

298
MCQeasy

What is the key advantage of multi-head attention over single-head attention in transformer models?

A.It eliminates the need for positional encoding
B.It reduces the total number of parameters
C.It allows the model to focus on different parts of the sequence simultaneously from different representation subspaces
D.It makes the model non-autoregressive
AnswerC

Each head learns different attention patterns, improving model capacity.

Why this answer

Multi-head attention allows the model to attend to information from different representation subspaces at different positions, capturing a richer understanding.

299
MCQeasy

A startup is building a chatbot for customer support using OCI Generative AI Service. The chatbot needs to answer queries about product features based on a knowledge base of product documentation. Which configuration is most appropriate for this use case?

A.Use the Summarization task type to generate concise answers from the documentation.
B.Use a Cohere Command model with the knowledge base as context in a prompt, and enable retrieval-augmented generation (RAG) via OCI Generative AI Agents.
C.Fine-tune a Llama 2 70B model on the product documentation to create a custom model.
D.Use the Code Generation model to produce SQL queries that retrieve answers from a database.
AnswerB

This approach uses a foundation model with RAG to ground responses in the knowledge base, which is ideal for question answering.

Why this answer

OCI Generative AI Agents with retrieval-augmented generation (RAG) allows the chatbot to dynamically retrieve relevant chunks from the product documentation knowledge base and inject them as context into a Cohere Command model prompt. This approach ensures answers are grounded in the latest documentation without requiring fine-tuning, and it scales efficiently as the knowledge base grows.

Exam trap

Oracle often tests the distinction between task-specific models (summarization, code generation) and the RAG architecture, leading candidates to mistakenly choose a simpler task type like summarization instead of recognizing the need for retrieval-augmented generation.

How to eliminate wrong answers

Option A is wrong because the Summarization task type is designed to condense a given text into a shorter summary, not to answer specific queries by retrieving and reasoning over a knowledge base; it lacks the retrieval component needed for question answering. Option C is wrong because fine-tuning a Llama 2 70B model on product documentation would be computationally expensive, requires significant labeled data, and does not easily accommodate updates to the documentation without retraining, making it impractical for a dynamic knowledge base. Option D is wrong because Code Generation models are specialized for generating code (e.g., SQL, Python), not for answering natural language questions from a knowledge base; using SQL queries would require a structured database schema, which is not the case for unstructured product documentation.

300
MCQeasy

A retail company wants to generate product descriptions from attribute data. They have no prior AI experience. Which approach is most appropriate?

A.Use the Cohere Command model with carefully crafted prompts.
B.Train a custom model from scratch.
C.Fine-tune a model on a synthetic dataset.
D.Use the Cohere Embed model to generate embeddings and then decode.
AnswerA

Cohere Command can generate descriptions directly with simple prompts, requiring no additional training.

Why this answer

The Cohere Command model is specifically designed for text generation tasks like creating product descriptions from attribute data. With no prior AI experience, using a pre-trained model with carefully crafted prompts is the most efficient and accessible approach, requiring no custom training or complex infrastructure.

Exam trap

OCI GenAI often tests the distinction between generative models (like Cohere Command) and embedding models (like Cohere Embed), leading candidates to mistakenly choose the Embed model for text generation tasks when it is only designed for semantic similarity and retrieval.

How to eliminate wrong answers

Option B is wrong because training a custom model from scratch requires extensive AI expertise, large labeled datasets, and significant computational resources, which is impractical for a company with no prior AI experience. Option C is wrong because fine-tuning a model on a synthetic dataset still requires AI knowledge to generate realistic synthetic data and manage the fine-tuning process, and synthetic data may not capture real-world nuances, leading to poor generalization. Option D is wrong because the Cohere Embed model generates embeddings (vector representations) for semantic similarity or search, not text generation; decoding embeddings back into coherent product descriptions is not a standard or supported capability of the Embed model.

Page 3

Page 4 of 11

Page 5

All pages