1Z0-1127-25 Practice Exam 51 — 20 Questions

Question 1

A data scientist fine-tuned a base model in OCI Generative AI on a proprietary dataset and now wants to serve inference. They must keep unit costs predictable and avoid the operational burden of provisioning GPU infrastructure themselves. Which deployment choice should they make?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Use on-demand model hosting for the fine-tuned model and call it through the inference endpoint using its OCID.. On-demand hosting serves a fine-tuned model through the managed inference endpoint without the customer managing GPU infrastructure, and its consumption-based billing keeps unit costs predictable. Referencing the model by OCID lets applications switch between base and fine-tuned models with minimal change. Dedicated AI clusters offer isolation and reserved capacity but shift scaling and capacity planning back to the customer.

Answer

Create a dedicated AI cluster with the fine-tuned model and manage the cluster's scaling manually.

Answer

Deploy the fine-tuned model to a dedicated AI cluster that auto-scales with a minimum of one node and a maximum of ten nodes.

Answer

Export the fine-tuned model artifacts and run them on a self-managed GPU instance in a compute cluster.

Question 2

A solutions architect must let an application call OCI Generative AI inference endpoints without embedding long-lived user credentials in source code. The application runs on OCI Compute instances, and the security team wants authentication handled through OCI Identity and Access Management with automatic credential rotation. Which approach meets these requirements?

Accepted Answer

Create an IAM policy that grants the Generative AI family permissions to a dynamic group matching the compute instances, and configure the SDK to use instance principals.. Instance principals give compute instances a platform-managed identity with automatically rotated certificates, so no secret needs to live in code. Authorizing a dynamic group that matches those instances through an IAM policy scoped to generative-ai-family permissions completes the least-privilege, rotation-friendly design. Static API keys, customer secret keys, or anonymous access all fail the stated security constraints.

Answer

Create an API signing key for an IAM user, embed the private key in the application configuration file, and rotate it quarterly with a script.

Answer

Generate a customer secret key for a service user, store it in the instance metadata, and read it at runtime from the metadata service.

Answer

Configure the OCI Generative AI endpoint to accept anonymous calls by disabling authorization on the inference route.

Question 3

A developer is building a summarization feature that calls the OCI Generative AI chat API. They want the model to produce short, factual summaries and avoid inventing details that are not present in the source text. Which parameter setting best supports this goal?

Accepted Answer

Set temperature to a low value such as 0.1 and craft a prompt that instructs the model to summarize only using information in the provided text.. Faithful summarization is best encouraged by combining a low temperature with a prompt that explicitly limits the model to information in the source. Low temperature reduces random sampling variance, and the prompt constrains the model's scope. Parameters such as top_k and max_tokens control diversity and length limits but do not themselves enforce factual grounding.

Answer

Set temperature to a high value such as 0.9 and leave top_p at its default.

Answer

Set top_k to a large value so more candidate tokens are considered at each step.

Answer

Increase max_tokens substantially so the model has more room to explain each summary point.

Question 4

A developer is building an OCI Generative AI application that streams generated text to end users. They call the OCI Generative AI inference endpoint with the appropriate chat request and need to process the response incrementally as tokens arrive rather than waiting for the full completion. Which capability of the OCI Generative AI inference API should they use?

Accepted Answer

Set the streaming flag on the chat request so the service returns a server-sent event stream of partial responses.. The OCI Generative AI chat inference API offers a streaming mode that emits partial responses as the model generates tokens, which is exactly what an application needs to display output incrementally. Other chat parameters such as maxTokens shape the completion but do not change how it is delivered, and unrelated operations like summarizeText or dedicated cluster hosting do not provide token-level streaming.

Answer

Increase the maxTokens parameter so the entire completion is buffered and returned faster in a single response.

Answer

Enable a dedicated AI cluster endpoint and poll the inference job status until the full text is available.

Answer

Call the summarizeText operation repeatedly with a growing conversation buffer to emulate incremental output.

Question 5

A team is evaluating OCI Generative AI for a customer service chatbot. They need to ensure the model's responses are factually accurate and do not include harmful content. Which TWO techniques should they implement to mitigate hallucinations and toxic outputs? (Choose two.)

Accepted Answer

Implement content moderation filters on both input and output. To mitigate hallucinations and toxic outputs, the team should ground responses in verified documents using RAG and apply content moderation filters to both inputs and outputs. These techniques directly address factual accuracy and harmful content. Other options like increasing temperature or using unverified data would worsen the issues.

Answer

Fine-tune the model on a large dataset of unverified social media posts

Answer

Set max tokens to a very low value to limit response length

Answer

Increase the temperature to make responses more creative

Question 6

A healthcare organization is deploying a custom model in OCI Generative AI Service. They need to ensure that the model's responses are compliant with HIPAA and do not contain protected health information (PHI). They also want to monitor for any attempts to extract PHI via adversarial prompts. Which TWO OCI Generative AI features should they configure? (Choose two.)

Accepted Answer

Audit logging of all inference requests and responses. To meet HIPAA requirements, the organization needs both a runtime content moderation layer to prevent PHI exposure and a monitoring mechanism to detect adversarial attempts. Guardrails with custom PHI detection enforces the policy, while audit logging provides visibility and accountability. Together, they form a defense-in-depth strategy for protecting sensitive health information.

Answer

Deploying the model in a private subnet with no internet access

Answer

Enabling automatic model retraining based on user feedback

Answer

Fine-tuning the model on a dataset of de-identified medical records

Question 7

An application team needs to send prompts to a Cohere model hosted in OCI Generative AI and receive completions. They want to authenticate without embedding long-lived credentials in code, and they are running on an OCI Compute instance. Which authentication approach should they use?

Accepted Answer

Instance principals with an IAM policy granting the dynamic group access to the generative AI family. For workloads on OCI Compute, instance principals are the recommended way to authenticate to OCI services without storing credentials. A dynamic group plus an IAM policy scoped to the generative AI family authorizes the instance to call the inference APIs. API keys, pre-authenticated requests, and basic auth are either insecure for this purpose or unsupported by the service.

Answer

Basic authentication with a username and password over HTTPS

Answer

A user's API key stored in an environment variable on the instance

Answer

Pre-authenticated requests (PARs) generated for the generative AI endpoint

Question 8

A team has created a dedicated AI cluster in OCI Generative AI and imported a fine-tuned model. Before the application can send inference requests to that model, the team must perform one more step. What should they do?

Accepted Answer

Create an endpoint on the dedicated AI cluster and associate the fine-tuned model with that endpoint.. A dedicated AI cluster supplies the hosting capacity, but the model becomes callable only when an endpoint is created on that cluster and bound to the imported fine-tuned model. Marketplace publication, retraining, and notebook attachment are unrelated to exposing an inference target, so creating the endpoint is the required next step.

Answer

Publish the fine-tuned model to the OCI Marketplace so it becomes globally addressable.

Answer

Attach the model to a data science notebook session and call it through the notebook kernel.

Answer

Convert the fine-tuned model into a generic base model by re-running the fine-tuning job with zero training examples.

Question 9

A team fine-tunes a base model in OCI Generative AI and hosts the resulting custom model on a dedicated AI cluster. During load testing they observe that concurrent inference requests are being rejected with throttling errors even though average utilization looks low. Which factor should they investigate first?

Accepted Answer

The dedicated AI cluster's hosting capacity, such as the number and size of compute units allocated for serving. Throttling during concurrent inference on a dedicated AI cluster is a capacity signal. The hosting capacity of the cluster, defined by the compute units provisioned for the endpoint, sets the concurrent request ceiling. Training epochs, sampling parameters, and unrelated embedding models do not affect that ceiling, so the cluster's serving capacity is the correct first place to investigate.

Answer

The temperature and top-p values configured in the inference request

Answer

The embedding model version used elsewhere in the application

Answer

The number of fine-tuning epochs used during training

Question 10

A RAG application on OCI Generative AI uses a vector store with 1 million chunks. Users report that answers are correct but slow. Profiling shows the vector search step dominates latency. You must reduce retrieval latency while preserving recall. Which approach should you take?

Accepted Answer

Switch from exact k-nearest-neighbor search to an approximate nearest-neighbor index such as HNSW or IVF.. For large vector stores, exact nearest-neighbor search scans all vectors and becomes the bottleneck. Approximate nearest-neighbor indexes such as HNSW or IVF reduce the number of distance computations by navigating a graph or partitioning the space, cutting latency dramatically while keeping recall high when parameters are tuned. Other options either harm recall, help only a subset of queries, or fail to change the algorithmic cost.

Answer

Reduce the embedding dimensionality by truncating vectors to the first 128 dimensions.

Answer

Move the vector store to a larger compute shape with more memory and CPU.

Answer

Cache the LLM responses for repeated queries and return the cached answer without retrieval.

Question 11

A support team is building a RAG assistant on OCI Generative AI. They want the assistant to answer questions using only the company's internal policy documents stored in an OCI Object Storage bucket. They need to convert those documents into vectors and store them in a vector database for similarity search. Which OCI service should they use to generate the vector representations of the document chunks?

Accepted Answer

OCI Generative AI embedding models, such as Cohere Embed, invoked through the Generative AI inference endpoint.. Generating vector representations from text requires an embedding model, and OCI Generative AI exposes embedding models such as Cohere Embed through its inference endpoint. The other services handle transcription, image analysis, or metadata cataloging, none of which produce text embeddings. Keeping embeddings inside OCI Generative AI also aligns IAM and networking with the generation step of the RAG pipeline.

Answer

OCI Speech to transcribe each policy document and emit vector coordinates for indexing.

Answer

OCI Vision to extract features from document images and store them as embeddings.

Answer

OCI Data Catalog to profile the documents and generate semantic hashes for similarity comparison.

Question 12

A solutions architect is designing a chatbot that must answer questions about internal HR policy documents stored in an OCI Object Storage bucket. The documents change weekly, and the team wants to avoid retraining the foundation model each time content changes. Which OCI Generative AI capability should be used to ground the model responses in the current documents?

Accepted Answer

Retrieval-augmented generation (RAG) using a vector store in the OCI Generative AI Agents feature. RAG with OCI Generative AI Agents lets the application retrieve relevant passages from a vector store built from the Object Storage documents and inject them into the prompt. This grounds answers in current content, avoids retraining, and handles weekly changes by re-indexing the knowledge base. The other options either retrain the model, ignore retrieval, or only change compute capacity.

Answer

Deploying the model on a dedicated AI cluster and enabling autoscaling for peak load

Answer

Fine-tuning the base model with a new dataset each week using the OCI Generative AI fine-tuning API

Answer

Increasing the model's context length to the maximum supported tokens so all documents fit in one prompt

Question 13

A company has deployed a fine-tuned model on an OCI Generative AI dedicated AI cluster. They want to programmatically invoke the model's endpoint from a Python application running on an OCI Compute instance. The application must authenticate without embedding long-lived credentials. Which OCI authentication method should they use?

Accepted Answer

Instance principals. Instance principals are designed for OCI Compute instances to authenticate to OCI services without long-lived credentials. By adding the instance to a dynamic group and granting that group permission to invoke the Generative AI endpoint, the application can obtain short-lived tokens automatically. This meets the security requirement and is the standard OCI pattern for compute-based workloads accessing OCI APIs.

Answer

API key authentication with a user's private key stored in the instance's boot volume

Answer

Resource principals with a dynamic group scoped to the Generative AI service

Answer

OAuth 2.0 client credentials flow with a confidential application in OCI IAM

Question 14

A solutions architect is deploying a custom model on an OCI Generative AI dedicated AI cluster. The model requires 4 GPUs to serve inference requests. The architect wants to ensure the endpoint can scale automatically based on incoming request volume while minimizing idle GPU cost. Which deployment configuration should the architect use?

Accepted Answer

Create a dedicated AI cluster with autoscaling enabled, specifying a minimum of 1 and maximum of 4 replica units.. Autoscaling in OCI Generative AI dedicated AI clusters adjusts the number of replica units based on workload. Setting a minimum of one unit ensures availability, while a maximum of four units allows scaling during peak demand. This configuration directly addresses the need for automatic scaling and cost reduction during idle periods, making it the correct deployment strategy.

Answer

Create multiple dedicated AI clusters, each with one GPU unit, and use a load balancer to distribute requests across them.

Answer

Create a dedicated AI cluster with a fixed size of 4 GPU units and enable load balancing across all units.

Answer

Deploy the model on OCI Data Science with a GPU shape and configure a scheduled scaling policy based on CPU utilization.

Question 15

A solutions architect is designing a retrieval-augmented generation (RAG) workflow using OCI Generative AI embeddings and a generation model. The knowledge base changes frequently and answers must cite the most current documents. Which two practices should the architect include? (Choose two.)

Accepted Answer

Pass the retrieved document chunks into the prompt as grounding context and instruct the model to answer only from that context. A current-answer RAG pipeline needs fresh retrieval data and grounded generation. Re-embedding and upserting changed documents keeps the vector index accurate, while injecting retrieved chunks into the prompt and constraining the model to that context produces answers that reflect the latest content. Temperature changes, passive storage, and constant fine-tuning do not deliver recency.

Answer

Store the original documents only in Object Storage and rely on the model's pretrained knowledge for answers

Answer

Increase the generation model's temperature on every request to force it to prefer newer documents

Answer

Fine-tune the generation model on the entire knowledge base after every document change

Question 16

A team is building a document Q&A service on OCI Generative AI. They embed a large corpus with the embedding model and store vectors in a vector database. During testing, users complain that answers to narrow questions are vague and sometimes cite unrelated sections. The team wants to improve retrieval precision without re-embedding the entire corpus. Which action should they take first?

Accepted Answer

Apply metadata filters and a reranking step over the retrieved candidates, and tune the number of results returned.. Vague answers with unrelated citations point to retrieval, not generation. Adding metadata filters, introducing a reranker over the candidate set, and tuning how many chunks are passed to the model improve precision at query time without touching the stored embeddings. Re-embedding, larger chunks, or bigger generation models do not address the root cause and some would make it worse.

Answer

Increase the chunk size for all documents so each vector represents a broader span of text.

Answer

Lower the embedding model's output dimensionality to speed up similarity search and reduce noise.

Answer

Switch the generation model to a larger parameter count so the LLM can reason over noisier context.

Question 17

A financial services firm runs a RAG chatbot on OCI Generative AI. The knowledge base contains both public marketing content and confidential client agreements in the same OCI OpenSearch vector index. Auditors require that users in the retail division never receive answers grounded in confidential client agreements, even if those chunks are semantically closest to the query. Which approach best enforces this at retrieval time?

Accepted Answer

Store a classification field on each chunk and apply a metadata filter in the OpenSearch vector query so only chunks the user's division is authorized to see are eligible for retrieval.. Access control must be enforced where retrieval happens. Tagging chunks with classification metadata and filtering the vector query ensures unauthorized documents are excluded before ranking, so they never reach the prompt. Application-side conventions or prompt instructions leave room for error and do not provide the deterministic guarantee auditors require for confidential client data.

Answer

Add a system prompt instruction telling the model to ignore confidential client agreements when answering retail users.

Answer

Create a separate OpenSearch index for confidential agreements and rely on the application to query only the public index for retail users.

Answer

Raise the similarity threshold so only the highest-scoring chunks are returned, reducing the chance a confidential chunk is included.

Question 18

A platform team is deploying a custom fine-tuned model on an OCI Generative AI dedicated AI cluster and wants the endpoint to serve predictable production traffic without manual intervention when a host fails. Which deployment configuration should they choose?

Accepted Answer

Create a dedicated AI cluster with multiple compute units per model endpoint and enable host failover across the cluster.. A dedicated AI cluster with multiple compute units per endpoint is the production-grade option: it gives isolated, reserved capacity and lets the service move traffic away from a failed host automatically. On-demand capacity, raw compute instances, and Object Storage cannot deliver that combination of managed hosting and resilient, predictable throughput.

Answer

Host the model on a single compute instance in a private subnet and attach a public load balancer in front of it.

Answer

Deploy the model on an on-demand inference endpoint and rely on the shared service capacity to absorb host failures.

Answer

Import the model into OCI Object Storage and let the Generative AI service pull it on each inference request.

Question 19

A platform team is preparing to host a custom fine-tuned model on an OCI Generative AI dedicated AI cluster. They need to expose the model through a private endpoint that is reachable only from application subnets inside their VCN. Which two actions are required to meet these requirements? (Choose two.)

Accepted Answer

Create the dedicated AI cluster with a private endpoint and select the target VCN and subnet during cluster creation. A private dedicated AI cluster endpoint must be created with the target VCN and subnet specified, and the VCN's security rules must allow application subnets to reach the cluster port. Together these place the endpoint inside the customer network and restrict access to authorized subnets. Public exposure options and service gateways do not satisfy the private-only connectivity requirement.

Answer

Attach an internet gateway to the VCN and publish the endpoint through a public load balancer

Answer

Enable a NAT gateway on the application subnet so outbound calls to the model endpoint resolve through the public internet

Answer

Create a service gateway so the application subnet can reach the OCI Generative AI service without traversing the internet

Question 20

A team deploys a RAG application on OCI Generative AI. They observe that when a user asks a follow-up question such as 'and what about the second option?', retrieval returns chunks unrelated to the original topic, and the model answers incorrectly. The application currently embeds only the raw follow-up text and searches the vector index with it. Which change best fixes the retrieval failure?

Accepted Answer

Rewrite the follow-up into a standalone query using the conversation history before generating the embedding for vector search.. A follow-up like 'and what about the second option?' lacks the subject needed for meaningful embedding. Rewriting it into a standalone query using conversation history restores the referent, so the query vector aligns with the correct chunks in the index. The other options either fail to add the missing context or degrade the user experience without fixing retrieval.

Answer

Increase the embedding model's output dimension so the query vector captures more of the conversational context.

Answer

Disable conversation history and require users to restate the full question each time so every query embeds cleanly.

Answer

Store the entire conversation transcript as a single chunk in the vector index and retrieve it for every follow-up question.