50 Oracle Cloud Infrastructure 2025 Generative AI Professional Practice Questions: Question Bank 2025

A product team wants to add a "summarize this page" feature to their internal knowledge portal. They want a fully managed OCI capability and do not want to deploy or manage model servers. Which OCI option best fits?

You are designing a Retrieval-Augmented Generation (RAG) solution to answer questions using internal policy PDFs. What is the primary purpose of the retrieval step in RAG?

A developer is calling an OCI Generative AI endpoint from an OCI compute instance. Which authentication approach is a recommended best practice to avoid long-lived user API keys on the instance?

A team is experiencing inconsistent answers from an LLM when asking the same question multiple times. They want more deterministic output. Which change most directly improves determinism?

A bank wants a chatbot that answers customer questions using only approved product documentation. The security team is concerned about hallucinations and wants responses to cite sources. Which architecture best addresses this requirement?

You must expose a generative AI capability to multiple internal applications. Security requires that only calls from a specific VCN are allowed, and the service should not be publicly reachable from the internet. Which OCI design is most appropriate?

A RAG pipeline is returning irrelevant context, leading to wrong answers. Initial investigation shows many PDFs contain repeated headers/footers and page numbers. What is the most effective first remediation step?

An application sends prompts that sometimes exceed the model’s context window, causing failures or truncated responses. Which design pattern is most appropriate to handle this reliably?

A healthcare organization must prevent accidental disclosure of PHI when using an LLM to draft responses. They want guardrails that reduce the chance of sending sensitive data to the model and also prevent sensitive data from appearing in outputs. Which combined approach is the strongest?

A development team wants to evaluate multiple prompt templates and model configurations for a customer support assistant. They need a repeatable, measurable process that compares response quality and safety before production rollout. What is the best practice approach?

A team wants to reduce hallucinations in a chatbot that answers questions strictly from an internal policy manual. They already have the manual stored as documents and want the model to cite content from it. Which approach is MOST appropriate?

You are configuring access for developers to call the OCI Generative AI service from a compute instance using instance principals. What is the BEST practice to control access?

A developer is building a chat application that maintains conversation context across multiple user messages. What technique should they use to preserve context while controlling token usage?

A bank is implementing a RAG solution on OCI. They want strong separation so that users can only retrieve passages from documents they are authorized to see. Which design is MOST appropriate?

An application calling OCI Generative AI intermittently fails with authorization errors, but only when deployed to a new subnet. The same code works in the old subnet. Which is the MOST likely root cause?

A team uses OCI Generative AI to generate customer support responses. Compliance requires that customer PII not be sent to the model. Which is the BEST approach to meet this requirement while preserving response usefulness?

You are optimizing a summarization workflow using OCI Generative AI for long legal documents. The output sometimes misses key sections. Which change is MOST likely to improve coverage without significantly increasing hallucinations?

A developer wants to implement function/tool calling so the LLM can retrieve an order status from an internal API. Which design is MOST appropriate to reduce security risk from prompt injection?

A regulated enterprise must demonstrate that only approved OCI Generative AI models are used in production and that changes are auditable. Which combination best supports this requirement?

A team built a RAG pipeline, but the model frequently answers with plausible content that is not present in retrieved passages. They confirm retrieval is returning relevant chunks. What is the MOST effective next step to improve faithfulness?

A product manager asks what "temperature" controls when generating text from an LLM. Which statement best describes temperature?

Your team is building an OCI-based GenAI assistant and wants to reduce hallucinations when answering questions that must be grounded in internal policy documents. Which approach is most appropriate?

A developer wants OCI to automatically rotate the customer-managed key used to encrypt data at rest for an application that stores LLM prompts and responses. Which OCI service capability should they use?

A team is troubleshooting inconsistent results from the same prompt sent multiple times to an LLM endpoint. They need more repeatable outputs for regression testing. Which change is most likely to improve determinism?

You must let multiple internal applications call OCI Generative AI while enforcing least privilege. Each application should only be able to invoke a specific model endpoint and nothing else. Which design best meets this requirement?

A solution uses RAG with a vector store. Users complain that answers cite irrelevant sections. You confirm the correct documents are indexed. Which tuning action is most likely to improve retrieval relevance before changing the LLM?

An application calls an LLM endpoint and intermittently receives HTTP 429 (Too Many Requests). The team wants a resilient fix that protects the endpoint while maintaining user experience. What is the best approach?

A regulated enterprise wants to ensure prompts containing sensitive data are not exfiltrated and that only approved egress destinations are allowed from the GenAI integration components. Which OCI architecture best supports this goal?

A team wants an LLM-based assistant to follow strict corporate rules (tone, refusal behavior, citation requirements) across multiple apps. They also need to keep application prompts small and consistent. What is the best pattern to achieve this?

A security review finds that an internal GenAI chatbot can be manipulated with prompt injection to ignore policy and reveal confidential snippets included in retrieved context. Which mitigation is most effective at the application level?

You are building a retrieval-augmented generation (RAG) assistant. During testing, the model sometimes answers with plausible but incorrect details even when the retrieved passages do not support them. Which change most directly improves groundedness in the response?

A developer wants to call OCI Generative AI from an application without embedding long-lived user credentials. Which approach is the recommended best practice for authentication from OCI compute-based workloads?

Your team needs the model to produce responses in a strict JSON schema (keys must always be present). Which prompting technique is most appropriate?

A team wants to minimize accidental exposure of sensitive data in prompts sent to OCI Generative AI. What is the most effective first step?

You deployed a RAG chatbot using embeddings. After switching to a different embedding model, search quality dropped significantly even though documents and queries appear similar. What is the most likely cause?

A product team wants to evaluate two prompt templates for the same use case and decide which produces fewer policy-violating outputs. Which approach best supports a controlled comparison?

An OCI Function calls the Generative AI service but intermittently fails with authorization errors after a refactor. The function runs in a private subnet and uses Resource Principals. Which is the most likely misconfiguration?

You need to build a multi-turn assistant where each request should remember prior user messages, but token usage must remain bounded to avoid context overflow. Which design is most appropriate?

A company must ensure that only approved project teams can invoke specific OCI Generative AI models, and all access must be auditable by compartment. Which design best meets this requirement?

You run an enterprise RAG solution where documents are updated frequently. Users complain that the assistant still cites old versions even after updates. What architecture change most directly addresses this while keeping retrieval fast?

Your team wants to reduce hallucinations in a customer support RAG assistant by ensuring the model answers ONLY from retrieved content and otherwise responds that it does not know. Which prompting approach best enforces this behavior?

A developer integrates OCI Generative AI into an internal tool and needs deterministic outputs for regression testing. Which parameter setting most directly supports this goal?

You are building an OCI Generative AI chat endpoint for multiple business units. Security requires that only a specific group can invoke the model. Which OCI capability is the MOST appropriate way to enforce this access control?

A RAG pipeline retrieves too many irrelevant chunks from a large document corpus, causing slow responses and noisy context. Without changing the model, what is the BEST first improvement to retrieval quality?

You must log prompts and responses for auditing but the data may contain PII. What is the BEST practice to meet governance requirements while minimizing exposure risk?

A team uses function/tool calling so the model can retrieve order status from an internal API. In testing, the model sometimes attempts to call the tool with malformed JSON arguments. What is the MOST effective way to reduce malformed tool calls?

An application must keep its inference traffic off the public internet and call OCI Generative AI from a private network. Which architecture pattern BEST meets this requirement?

You observe that a summarization workflow occasionally returns partial summaries truncated mid-sentence. The prompts and input sizes vary widely. What is the MOST likely cause and the best corrective action?

A regulated enterprise must deploy a RAG system where the vector store is updated daily. Security requires demonstrating that no unauthorized documents can influence model answers. Which design provides the STRONGEST control for this requirement?

A multi-step agent uses a model to plan, call tools, and then generate a final answer. In production, you see intermittent failures where the agent loops between planning and tool calls, consuming excessive tokens. What is the BEST mitigation strategy?