Agentic AI for Enterprises (OCI Enterprise AI) practice questions

Worked examples for Agentic AI for Enterprises (OCI Enterprise AI)

1. Question 1

   A platform team is writing the operational readiness checklist for an **OCI Generative AI Agents** assistant that is about to serve thousands of employees. During the pilot the team judged the assistant purely on whether its answers were correct. A senior architect argues that answer quality alone is an insufficient bar for a production enterprise deployment, and that the checklist must cover concerns the pilot never had to face. Which TWO concerns become materially more important when this agent moves from pilot to production, as opposed to remaining a small internal demo?

   1. A. Keeping the knowledge base synchronized with the underlying enterprise data sources, so that answers reflect current content instead of a stale snapshot ingested at build timeCorrect answer

      Correct. A production agent answers continuously against documents that keep changing, so the ingestion of source data into the knowledge base must be maintained on an ongoing basis; a pilot could tolerate a one-time snapshot because it was only exercised briefly.

   2. B. Operating and scaling the agent endpoint that serves user sessions, including its cost and latency behaviour under sustained real-world loadCorrect answer

      Correct. In production the agent is consumed through a hosted endpoint that must sustain many concurrent sessions; throughput, latency, and the cost of that traffic are real operational obligations that a handful of demo conversations never surfaces.

   3. C. Retraining the underlying large language model on the company's documents each time those documents change, because retrieval only works on data the model was trained on

      Rests on the misconception that retrieval-augmented generation requires retraining the model. The agent retrieves from a knowledge base at request time; keeping content current is an ingestion concern, and no model retraining is involved.

   4. D. Nothing new — a deployed agent behaves exactly like the pilot, so the checklist only needs to re-measure answer accuracy at a larger sample size

      Embodies the misconception that a production enterprise agent is no different from a demo. Scale, real users, live data, and audit obligations introduce operational and governance concerns that accuracy sampling does not measure at all.

   5. E. Disabling response monitoring once the pilot demonstrates good answers, since a trusted model no longer needs observation and logging only adds overhead

      Claims observability is unnecessary because the model is trusted. Production behaviour drifts as data, users, and question mixes change, and enterprise deployments need traceable records for audit — removing monitoring is the opposite of production readiness.

   Explanation

   Moving an agent into production changes the problem from 'does it answer well in a demo' to 'can it be operated'. Because OCI Generative AI Agents grounds answers in a knowledge base built by ingesting enterprise data sources, that ingestion has to be kept current or the assistant confidently serves stale content; and because users reach the agent through a hosted agent endpoint, sustained concurrency, latency, and cost become genuine operational obligations. The idea that changed documents require retraining the model misreads retrieval-augmented generation, which consults the knowledge base at request time. Assuming a deployment behaves exactly like the pilot ignores scale, live data, and audit duties, and switching off monitoring because the model performed well removes precisely the evidence a regulated enterprise needs.

2. Question 2

   A production assistant built on **OCI Generative AI Agents** answers employee questions by retrieving from a knowledge base built over the company's internal documents. A subject-matter reviewer flags one answer as unsupported by any approved document. Before escalating, the operations team must determine, for that specific response, which knowledge-base content the agent actually grounded the answer on — and they need this to be routine practice, not a one-off forensic exercise, because the deployment is subject to internal audit. Which property of a retrieval-augmented response from OCI Generative AI Agents most directly supports this?

   1. A. Enable sessions on the endpoint so the agent retains conversation history across turns

      Confuses conversational memory with auditability. Sessions let the agent carry context from earlier turns of a chat so follow-up questions make sense; they say nothing about which document content grounded any particular answer.

   2. B. Responses are returned with citations that identify the source content in the knowledge base used to ground the answerCorrect answer

      Correct. Citations accompany an agent's retrieval-augmented response and point back to the knowledge-base content behind it, so a reviewer can check the answer against its actual grounding instead of judging the generated text on its own — the per-response evidence trail an audited enterprise deployment depends on.

   3. C. Re-ingest the data source with a different chunking configuration so that retrieval quality improves

      Treats a retrieval-tuning knob as a diagnostic mechanism. Re-ingesting may change future behavior, but it produces no evidence about the response under review — and without that evidence the team cannot even tell whether chunking was the cause.

   4. D. No investigation is required, because the underlying model is Oracle-managed and validated, so an unsupported answer necessarily means the source documents are wrong

      The misconception that a trusted, managed model removes the need for observability. A managed model can still retrieve a poorly matching passage or generate content that goes beyond it, so an enterprise deployment needs per-response evidence rather than an assumption of correctness.

   Explanation

   OCI Generative AI Agents grounds answers in a knowledge base and returns citations pointing to the source content used, which lets a reviewer trace a specific answer back to the material behind it (OCI Generative AI Agents overview). That traceability is what separates an enterprise deployment from a toy agent: an audited production assistant must be able to show, per response, where an assertion came from, whereas a prototype is judged only on whether the text reads well. Conversational sessions address continuity across turns, not evidence for a single answer; re-ingesting with different chunking changes future retrieval without documenting what already happened; and treating a managed model as beyond inspection ignores that grounding and generation can still go wrong in ways only per-response evidence exposes.

3. Question 3

   A compliance team runs an assistant built on OCI Generative AI Agents. The assistant is grounded on a **knowledge base** built from expense-policy documents that the team stores in OCI Object Storage. The policy documents are revised every quarter. After the latest revision, employees notice the assistant still answers with the *previous* quarter's per-diem rule, and cites the old document text. What is the correct way to make the grounded assistant reflect the revised policy?

   1. A. Fine-tune the underlying large language model on the revised policy documents, because grounding works by writing the enterprise content into the model's weights.

      This is the 'RAG retrains the model' misconception. Grounding through a knowledge base is retrieval at query time — the documents are chunked, embedded, and indexed outside the model; the model's weights are never modified by attaching or updating a knowledge base.

   2. B. Do nothing beyond publishing the new documents — the base model continuously learns from the company's Object Storage buckets and will pick up the revision on its own.

      This is the 'the base model already knows our data' misconception. A pretrained model has no awareness of a tenancy's private documents at all; enterprise content reaches an answer only because a configured data source was ingested and is retrieved at query time.

   3. C. Re-run ingestion for the knowledge base's data source so the revised documents are re-chunked, re-embedded, and indexed, after which retrieval returns the new passages to the model.Correct answer

      Correct. The knowledge base is only as current as its last ingestion: the agent retrieves from the indexed vectors, so refreshing the index is what changes the grounded answer. The model itself is untouched — it simply reasons over the newly retrieved passages.

   4. D. Clear the agent's session memory, since the previous policy text is being replayed out of the conversation history that memory keeps.

      This confuses memory with the vector store. Session memory holds the turns of a conversation so follow-up questions resolve in context; it is not where document content lives, so clearing it cannot replace a stale index.

   Explanation

   An agent grounded with a knowledge base answers by retrieving indexed passages from its ingested data sources and passing them to the model, which then composes the response — a retrieval-augmented generation pattern, not a training pattern. Therefore staleness in the answers is staleness in the index, and the fix is to ingest the revised source documents again so their embeddings replace the old ones. Changing model weights is neither necessary nor how the grounding path works; a pretrained model has no independent access to private tenancy documents; and conversation memory stores dialogue state, not the enterprise corpus, so clearing it addresses nothing here.

4. Question 4

   A product owner is reviewing a proposal to build an employee assistant on **OCI Generative AI Agents**, grounded on a knowledge base of internal documents and given a tool that calls the HR system's API. She pushes back: "We already have semantic search over that same indexed content. Employees can type a question and get back the matching passages. What does wrapping an agent around it actually buy us?" Which TWO statements correctly describe what a grounded agent adds over exposing the retrieval index as a search box? (Select TWO.)

   1. A. Once the documents are indexed, the large language model drops out of the request path entirely, and the agent simply returns the stored passage that scored highest.

      The 'vector store is the reasoner' misconception. Retrieval and generation are distinct steps: the store supplies candidate passages, and the model still reasons over them to produce the answer — which is the very thing a bare search box does not do.

   2. B. Semantic search hands back ranked passages that the employee must read and interpret, whereas a grounded agent passes those retrieved passages to the model, which synthesizes a direct answer to the question that was actually asked.Correct answer

      Correct. Grounding combines retrieval with generation: the knowledge base supplies the relevant enterprise context and the model reasons over it, turning matching passages into an answer rather than a reading assignment.

   3. C. Registering the HR system's API as a document in the knowledge base makes that API executable, so a single retrieval index removes any need for a separate tool.

      Confuses retrievable text with a callable capability. Indexing a description of an API only makes that description findable; taking an action against a live system requires the agent to invoke a tool, which is a different mechanism from retrieval.

   4. D. The agent chooses per request which attached capability fits — retrieving from the knowledge base for a policy question, calling the tool to act on the HR system — and carries session context across turns, so a multi-step request can be completed in one conversation.Correct answer

      Correct. An agent orchestrates: it plans which knowledge base or tool a given request needs and retains conversation state across turns, which a stateless search index cannot do.

   5. E. Attaching a knowledge base removes the burden of keeping content current, because the agent extrapolates the present policy from the older indexed versions.

      Assumes grounding repairs stale source data. A grounded answer is only as current as the ingested content; if superseded documents are what gets retrieved, the agent will ground on them, so keeping the knowledge base curated remains an ongoing responsibility.

   Explanation

   An agent on OCI Generative AI Agents does more than look content up: the knowledge base retrieves the enterprise context, the model reasons over that retrieved context to produce an answer, and the agent decides per request whether a question needs retrieval, a tool call against a live system, or both — while session state keeps a multi-turn request coherent. That rules out treating the index as the answerer with the model removed from the path, since retrieval and generation remain separate steps. It also rules out indexing an API description as a substitute for a tool, because retrievable text is findable but not executable, and it rules out expecting grounding to compensate for stale content, since the answer is only as current as what was ingested.

5. Question 5

   A support team builds an assistant on OCI Generative AI Agents. The assistant is given a **knowledge base backed by a vector store** built from the company's product manuals, and it is also configured to retain **conversation memory** across the turns of a chat session, so a user can ask "Does that model support 240V?" right after asking about a specific appliance. Which statement correctly distinguishes the role of the agent's conversation memory from the role of the vector store knowledge base?

   1. A. Memory carries the running state of the conversation so follow-up turns resolve against what was already said, while the vector store supplies passages retrieved from the ingested manuals that are relevant to the current question.Correct answer

      Correct division of labor described in the OCI Generative AI Agents overview: a knowledge base built over a vector store provides retrievable enterprise content, while session memory preserves conversational context so multi-turn questions such as "that model" stay resolvable.

   2. B. They are two names for the same mechanism: the embeddings created during a chat session become the agent's knowledge base, so configuring one automatically configures the other.

      Conflates memory with the vector store. Session memory is conversational state maintained across turns; the knowledge base is a separately ingested and indexed corpus of enterprise documents. Configuring one does not create the other.

   3. C. The vector store holds the running conversation history, while memory holds the embedded chunks of the product manuals that are retrieved at query time.

      Inverts the two roles. The vector store indexes the ingested document chunks for similarity retrieval; the conversational state is what memory carries, not the other way round.

   4. D. Once the manuals are ingested, neither is needed at query time, because ingestion updates the underlying model's weights with the manual content.

      Assumes ingestion fine-tunes the model. Ingesting documents into a knowledge base builds a retrieval index; it does not modify model weights, and retrieval plus conversational state are still required at query time.

   Explanation

   On OCI Generative AI Agents, grounding on business data comes from a knowledge base whose ingested content is indexed in a vector store and retrieved by similarity at query time, while conversational memory is the separate mechanism that keeps a multi-turn session coherent. Treating the two as one mechanism, or swapping which one holds documents versus dialogue state, misdescribes both. Nothing about ingesting documents changes the underlying model's parameters, so the claim that retrieval becomes unnecessary after ingestion is also wrong.

6. Question 6

   A team is promoting a prototype assistant built on **OCI Generative AI Agents** into production for a regulated business unit. The prototype answered questions well, but it ran with no controls beyond that. Before go-live, compliance requires two things: a reviewer must be able to see which enterprise source a given answer was drawn from, and the assistant must be prevented from returning unsafe content or exposing personal data in its responses. Which TWO capabilities of OCI Generative AI Agents directly address these enterprise requirements?

   1. A. Fine-tune the underlying large language model on the business unit's documents so that the compliance policies are baked into the model weights

      Misconception that governance is achieved by retraining the model. OCI Generative AI Agents grounds answers by retrieving from a knowledge base at request time (RAG), not by training policy into weights; weights also cannot be audited per answer, and a document removed from the source would still be reflected in a trained model.

   2. B. Enable guardrails on the agent endpoint, such as content moderation, personally identifiable information (PII) detection, and prompt-injection protectionCorrect answer

      Correct. Agent endpoints expose configurable guardrails — content moderation, PII detection, and prompt-injection protection — which is the service-level control for keeping unsafe or personal content out of production responses.

   3. C. Rely on the fact that the service is fully managed by Oracle, which removes the need for the deployment to apply its own response controls

      Misconception that a managed service transfers the customer's governance obligation to the provider. Oracle manages the infrastructure and models, but guardrails, access policies, and review of what the agent may say and see remain the deploying organization's configuration responsibility.

   4. D. Use the citations returned with a response, which identify the knowledge-base source content the answer was grounded inCorrect answer

      Correct. Responses carry citations back to the retrieved source content, which is what lets a reviewer trace an answer to the enterprise document behind it — the auditability property a prototype typically lacks.

   5. E. Treat the production agent exactly as the prototype was treated, since the same underlying model serves both and its behavior is unchanged

      Misconception that a deployed enterprise agent is no different from a demo. The same model in production faces real user data, real regulatory exposure, and real auditors, which is precisely why guardrails, citations, and access control are configured at deployment time.

   Explanation

   OCI Generative AI Agents is designed for enterprise deployment: agent endpoints carry configurable guardrails (content moderation, PII detection, prompt-injection protection) for what the agent may emit, and responses carry citations that tie an answer back to the knowledge-base content it was grounded in, giving auditors a traceable path from answer to source. Retraining the model on the corpus is not how the service grounds answers and produces nothing a reviewer can audit per response. Assuming a fully managed service absolves the deployment of its own controls confuses Oracle's infrastructure responsibility with the customer's governance responsibility, and carrying prototype practices into production ignores that the enterprise setting — regulated data, auditability, real users — is what imposes these constraints in the first place.

7. Question 7

   An enterprise deploys an OCI Generative AI Agents assistant that lets employees ask questions over an internal knowledge base. Sales, Legal, and HR documents are all ingested, but sales representatives must never see documents that belong to the Legal or HR teams. The team is deciding how to keep each group's questions restricted to the material that group is allowed to see. Which approach reflects the **data governance** expectation for an agent deployed in this enterprise setting?

   1. A. Scope access so each group's agent works only over the knowledge base content it is permitted to use, and enforce that access with OCI IAM policies on the agent and its resourcesCorrect answer

      Correct. OCI Generative AI Agents resources — agents, knowledge bases, and their data sources — are IAM-governed OCI resources in a compartment, so entitlement is enforced by policy and by which content each agent is pointed at, independent of what the model is asked to do.

   2. B. Add a sentence to the agent's instructions telling the model not to reveal Legal or HR content, and rely on the model to obey it

      Treats a natural-language instruction as an access-control mechanism. Prompt wording is a best-effort behavioural nudge, not an enforced boundary — the restricted documents are still retrievable, so a rephrased or adversarial question can surface them. Enterprise data governance requires enforcement outside the model.

   3. C. Skip access controls, because the agent only retrieves from documents the company already owns and therefore no data is exposed

      Confuses ownership of data with authorization to see it. Internally owned documents still carry per-group confidentiality; a deployed enterprise agent must respect need-to-know boundaries, not just the company perimeter.

   4. D. Treat governance as a prototype-only concern and rely on answer accuracy alone, since a correct answer cannot be a policy violation

      Assumes a deployed enterprise agent has no obligations beyond functional correctness. A factually accurate answer drawn from an HR file the asker may not see is precisely the violation governance exists to prevent, and governance concerns grow rather than disappear in production.

   Explanation

   In an enterprise deployment, an agent's access to data must be enforced by the platform rather than by the model's willingness to comply. OCI Generative AI Agents exposes agents, knowledge bases, and data sources as OCI resources living in compartments, so administrators restrict who can use which agent and what content each agent draws on through IAM policies — a control the model cannot talk its way around. Relying on a prompt-level instruction leaves the sensitive content retrievable, assuming that company-owned data needs no controls ignores need-to-know boundaries that exist inside the company, and treating accuracy as the only production concern misses that a correct answer delivered to an unauthorized reader is itself the breach.

8. Question 8

   An HR team is building an agent on OCI Generative AI Agents that must do two different kinds of work: answer employee questions from a large set of internal policy PDFs, and actually **file a time-off request** in the company's HR system of record when an employee asks for one. Which approach correctly assigns each of these responsibilities to the right agent capability?

   1. A. Register the HR system as an extra data source in the vector-store knowledge base; once its records are embedded, retrieving the right record is what files the request.

      This is the misconception that retrieval performs actions. A vector store supports read-only similarity search over ingested content; retrieving a record never writes to a downstream system. Creating a request requires a tool that invokes the HR system's API.

   2. B. Attach only the API tool; grounding on the policy PDFs happens automatically because ingesting documents into the vector store fine-tunes the underlying model on their content.

      This is the "RAG retrains the model" misconception. Ingesting and embedding documents builds a searchable index; it does not change the model's weights. Without retrieval at query time the model has no access to the policy content.

   3. C. Ground policy answers on a knowledge base that retrieves relevant passages from the policy documents, and give the agent a tool that calls the HR system's API to create the request.Correct answer

      Correct. OCI Generative AI Agents grounds answers by retrieving from a knowledge base over ingested enterprise content, and uses tools (such as function calling to an external endpoint) to take actions against enterprise systems — retrieval for knowledge, tools for actions.

   4. D. Attach only the knowledge base; no tool is needed, because once the model has read the relevant policy it can generate and execute the HR system's API call on its own.

      This confuses generating text about an action with performing it. The model can only produce output; invoking an external system requires a configured tool that the agent runtime actually executes.

   Explanation

   An enterprise agent grounds on business data by combining complementary capabilities: a knowledge base backed by a vector store makes enterprise documents retrievable so answers cite real policy content, and tools let the agent invoke enterprise APIs to take action. Retrieval is read-only and cannot write to a system of record, so loading HR records into the knowledge base does not file anything. Ingestion into a vector store builds an index rather than altering the model's weights, so it is not fine-tuning, and a model on its own can only produce text — executing an external call requires a tool the agent runtime invokes.