OCI Generative AI Service and Oracle Database 23ai practice questions

Worked examples for OCI Generative AI Service and Oracle Database 23ai

1. Question 1

   A team has finished fine-tuning a custom model in the OCI Generative AI service and now wants its application to send inference requests to that custom model. What must the team do to make the custom model callable?

   1. A. Nothing further — custom models become available automatically in the shared on-demand pool alongside the pretrained base models.

      Assumes custom models are served like base models on demand. A customer-specific model is served from the customer's own hosting capacity, not the shared pool.

   2. B. Create an endpoint for the custom model on a hosting dedicated AI cluster, and direct application requests to that endpoint.Correct answer

      Correct: a custom model is served by creating an endpoint on a hosting dedicated AI cluster, which is the address the application calls for inference.

   3. C. Export the fine-tuned model weights and deploy them onto an OCI Compute GPU instance running an inference server.

      Represents the 'you must host the model yourself' misconception. The managed service hosts custom models; weights are not exported for customer-run serving.

   4. D. Send requests to the fine-tuning cluster that produced the model, since that cluster continues to serve it after training completes.

      Conflates the two dedicated AI cluster types. A fine-tuning cluster trains the model; serving requires a separate hosting cluster and endpoint.

   Explanation

   Serving a fine-tuned custom model in the OCI Generative AI service means creating an endpoint on a hosting dedicated AI cluster and pointing the application at that endpoint. Custom models do not appear in the shared on-demand pool that fronts the pretrained base models, so no-action-required is wrong. Exporting weights to self-host on Compute contradicts the managed design of the service, and expecting the fine-tuning cluster to serve traffic conflates the training cluster type with the hosting cluster type.

2. Question 2

   A team is deciding which Oracle Database 23ai AI feature fits each of two requests: (1) "List the ten highest-value orders placed in Europe this month" against a normalized orders schema, and (2) "Find support tickets describing a problem like this customer's complaint" against a table of free-text ticket bodies. Which pairing of features to requests is correct?

   1. A. Select AI for the orders request and AI Vector Search for the support-ticket requestCorrect answer

      Correct. Select AI turns a natural-language question into SQL for structured, deterministic querying, while AI Vector Search performs similarity search over embeddings of unstructured text.

   2. B. Select AI for both requests, because natural-language translation to SQL covers structured aggregation and unstructured similarity equally well

      Represents the misconception that natural-language-to-SQL subsumes semantic search. Generated SQL matches literal predicates and cannot rank free-text tickets by conceptual similarity to a complaint.

   3. C. AI Vector Search for both requests, because embedding similarity can also compute exact aggregates and rankings over order values

      Represents the misconception that similarity search replaces deterministic SQL. Vector distance ranks by semantic closeness and cannot produce an exact top-ten-by-value aggregate.

   4. D. AI Vector Search for the orders request and Select AI for the support-ticket request

      Inverts the two features' roles — the classic swap of natural-language-to-SQL with semantic similarity search.

   Explanation

   The two features address different query shapes: natural-language-to-SQL translation is for structured questions whose answers come from deterministic SQL over schema columns — filters, joins, aggregates, and rankings — whereas similarity search over stored vector embeddings is for finding unstructured content that is conceptually close to an input, even with no shared wording. Applying either feature to both requests fails, because generated SQL cannot rank free text by meaning and vector distance cannot compute an exact top-N by monetary value. Reversing the assignments compounds both errors.

3. Question 3

   An HR department wants an assistant that answers employee questions from internal policy documents. The documents are revised weekly, and every answer must reflect the current revision. Which approach within the OCI Generative AI service best fits this requirement?

   1. A. Fine-tune a custom model on the policy documents and repeat the fine-tuning job after each weekly revision so the model stays current.

      Represents the 'fine-tuning is how you keep content fresh' misconception. Fine-tuning bakes knowledge into parameters and would require retraining on every revision, which retrieval avoids.

   2. B. Raise the model's temperature so it generates a wider range of answers and is therefore more likely to match the newest policy wording.

      Confuses a sampling parameter with knowledge freshness. Temperature controls randomness of the output, and cannot introduce content the model has no access to.

   3. C. Use a Generative AI Agent backed by a knowledge base built from the policy documents, so each answer is grounded in the currently indexed content.Correct answer

      Correct: Generative AI Agents ground responses in a knowledge base of enterprise content retrieved at query time, so refreshing the indexed documents updates the answers without retraining.

   4. D. Send requests to a base model with a large context window, since a larger context causes the model to permanently learn any documents it has processed.

      Represents the belief that context becomes permanent knowledge. Content in the context window influences only that request; it does not persist into the model.

   Explanation

   When the grounding content changes frequently, the retrieval-based approach fits: a Generative AI Agent answers from a knowledge base of enterprise documents retrieved at query time, so reindexing the revised policies is enough to keep answers current. Repeated fine-tuning bakes knowledge into model parameters and forces a retraining cycle for every revision. Temperature governs randomness in sampling rather than what the model knows, and a large context window influences only the request that carries the text — nothing processed in context is retained by the model afterwards.

4. Question 4

   A business analyst with no SQL training types "which three regions had the highest sales last quarter?" against an Oracle Database 23ai schema and receives a result set. Which statement best describes what Select AI did?

   1. A. It fine-tuned a large language model on the sales tables so the model could answer future questions from its own weights

      Reflects the misconception that Select AI trains or customizes a model on your data. Select AI performs inference-time translation of a question into SQL; it does not train, fine-tune, or absorb table contents into model weights.

   2. B. It used an LLM together with schema metadata to generate a SQL statement for the question and ran that SQL against the databaseCorrect answer

      Correct. Select AI augments the prompt with database metadata, has an LLM produce SQL, then executes it in the database and returns rows — natural-language querying of your own data.

   3. C. It embedded the question as a vector and returned the sales rows whose embeddings were most similar to it

      Conflates Select AI with AI Vector Search. Similarity search over embeddings retrieves semantically near items; it cannot compute an aggregate ranking such as "top three regions by sales", which requires generated SQL.

   4. D. It matched the question against a library of pre-written report queries shipped with the database and ran the closest one

      Assumes a canned-report catalog. Select AI generates SQL dynamically for the user's own schema; there is no built-in fixed query library that questions are matched to.

   Explanation

   Select AI is natural-language querying: the database sends the user's question plus relevant schema metadata to a configured LLM, receives SQL, and executes it — so users who do not write SQL can still query governed enterprise data in place. It is an inference-time translation step, not model training, so nothing about the sales tables is learned into weights. It is also distinct from embedding-based similarity retrieval, which finds semantically close items rather than computing aggregates, and it does not depend on any shipped catalog of canned queries.

5. Question 5

   A team has fine-tuned a custom model in OCI Generative AI and now needs to serve it with predictable, isolated inference capacity for a production application. Which statement most accurately describes the role of a dedicated AI cluster in this scenario?

   1. A. Dedicated AI clusters exist only to run fine-tuning jobs; once a custom model is trained, it can be served only through the shared on-demand inference endpoint.

      The misconception that dedicated clusters are a training-only construct. The service defines both fine-tuning clusters and hosting clusters, and a custom fine-tuned model must be placed on a hosting dedicated AI cluster to be served through an endpoint.

   2. B. A dedicated AI cluster is a set of GPU instances the customer provisions in their own tenancy, on which they install the model weights and inference server themselves.

      The misconception that OCI Generative AI is self-hosted infrastructure. It is a fully managed service: Oracle provisions and operates the cluster hardware and model serving; the customer never installs weights or an inference runtime.

   3. C. A dedicated AI cluster is managed, single-tenant compute allocated to the customer's workload — one cluster type sized for fine-tuning and another for hosting a model behind an endpoint — so the GPUs are not shared with other tenants.Correct answer

      Correct per the service overview: dedicated AI clusters are dedicated GPU compute for generative AI workloads, isolated from other GPUs/tenants, and come in fine-tuning and hosting flavors; a custom model is served by hosting it on a hosting cluster.

   4. D. A dedicated AI cluster is the capacity tier used by the prebuilt OCI AI Services (Language, Vision, Speech), and Generative AI simply reuses the same clusters.

      Confuses the Generative AI service with the prebuilt AI Services. The prebuilt services are consumed through their own task-specific managed APIs and do not expose or require dedicated AI clusters.

   Explanation

   Dedicated AI clusters are managed, single-tenant GPU compute inside the OCI Generative AI service, isolated from other customers' workloads, and the service distinguishes clusters sized for fine-tuning from clusters that host a model behind an inference endpoint — so a custom fine-tuned model is served by attaching it to a hosting cluster rather than to the shared on-demand endpoint. Treating clusters as training-only ignores the hosting cluster type; treating them as customer-provisioned GPUs contradicts the managed nature of the service, where Oracle operates the hardware and model serving; and equating them with prebuilt AI Services capacity conflates two different products, since Language, Vision, and Speech are consumed through fixed task APIs with no cluster concept. See the OCI Generative AI service overview.

6. Question 6

   A customer fine-tunes one of the pretrained foundation models in the OCI Generative AI service using its own labeled training data. What is the result of that fine-tuning operation?

   1. A. The shared pretrained base model is updated in place, so its behaviour changes for every tenancy that calls it afterwards.

      Represents the belief that fine-tuning mutates the shared base model. Fine-tuning yields a private custom model; the base model other customers call is unaffected.

   2. B. The training examples are stored and injected into every future prompt at inference time, which is what makes the model appear specialized.

      Confuses fine-tuning with few-shot prompting. Few-shot places examples in the prompt at inference time; fine-tuning adjusts model parameters ahead of inference.

   3. C. The training documents are indexed into a vector store that the model searches on each request to compose its answer.

      Confuses fine-tuning with retrieval-augmented generation. RAG retrieves passages at query time; fine-tuning changes the model itself rather than adding a search step.

   4. D. A separate custom model is created for the customer — a copy of the base model adapted to the customer's data — while the base model remains unchanged.Correct answer

      Correct: fine-tuning in the service produces a customer-specific custom model derived from the chosen base model, leaving the pretrained base model untouched for other users.

   Explanation

   Fine-tuning in the OCI Generative AI service takes a pretrained base model and a customer's own training data and produces a distinct custom model belonging to that customer; the pretrained base model itself is unchanged and other tenancies see no difference. Claiming the shared base model is updated misstates that isolation. The remaining choices swap fine-tuning for two other adaptation techniques: putting examples in the prompt is few-shot prompting, and indexing documents for lookup at query time is retrieval-augmented generation — neither changes model parameters the way fine-tuning does.

7. Question 7

   A regulated bank wants semantic search over customer records but its security policy forbids copying customer data outside the database of record. What is the principal architectural advantage of using Oracle Database 23ai AI Vector Search rather than a separate, standalone vector store?

   1. A. It removes the need for an embedding model, because the database derives semantic meaning directly from column data types

      Represents the misconception that vector search works without embeddings. Vectors must still be generated by an embedding model; the database stores and searches them, it does not infer meaning from data types.

   2. B. Embeddings are stored in a native VECTOR column alongside the business data, so similarity search and relational filtering happen in one system under one set of security and transactional controlsCorrect answer

      Correct. AI Vector Search's stated benefit is bringing AI to the data: vectors live with the relational rows, so a single SQL statement can combine similarity search with normal predicates, and existing database security, consistency, and backup apply.

   3. C. It guarantees that generated answers are factually correct, because the database validates LLM output before returning it

      Represents the misconception that grounding eliminates hallucination by guarantee. Vector search improves relevance of retrieved context but performs no correctness validation of model output.

   4. D. It converts the customer tables into an unstructured document collection, so relational schema and SQL joins are no longer used

      Represents the misconception that adopting vector search abandons the relational model. The VECTOR type is an additional column type; tables, schemas, and joins remain fully intact.

   Explanation

   AI Vector Search adds vectors as a native database data type so embeddings sit in the same tables as the business data, letting one SQL statement mix similarity search with ordinary relational predicates while inheriting the database's existing security, transactional consistency, and backup — the essence of bringing AI to the data instead of moving data to the AI. An external embedding model is still required to produce the vectors, so the feature does not eliminate that step. Grounding retrieval in enterprise data improves relevance but offers no guarantee of factual correctness in generated text. And the relational model is preserved: vector columns supplement tables rather than replacing schemas and joins.

8. Question 8

   A business analyst with no SQL skills needs to ask questions such as "what were total sales by region last quarter?" directly against tables in an Oracle Database 23ai schema. Which statement best describes what Select AI does to satisfy this request?

   1. A. It fine-tunes a large language model on the schema's data so the model can answer sales questions from its own trained weights

      Represents the misconception that Select AI trains or customizes a model. Select AI performs no training; it uses an LLM at query time to generate SQL, and the answer comes from executing that SQL against live data.

   2. B. It embeds every row of the sales tables as vectors and returns the rows whose embeddings are nearest to the question

      Confuses Select AI with AI Vector Search. Similarity search over embeddings retrieves semantically close content; it cannot compute an exact aggregate such as total sales by region.

   3. C. It uses an LLM together with schema metadata to translate the natural-language question into SQL, then runs that SQL in the database and returns the resultCorrect answer

      Correct. Select AI is the natural-language-to-SQL feature: the database supplies schema context to an LLM, the LLM generates SQL, and the database executes it so the answer comes from real data.

   4. D. It is a new SQL clause that improves optimizer plans for analytic queries and has no connection to language models

      Represents the misconception that 'Select AI' is merely a SQL performance keyword. Its entire purpose is LLM-driven natural-language querying, not plan optimization.

   Explanation

   Select AI lets users pose questions in natural language; the database passes the question plus relevant schema metadata to a configured large language model, which returns SQL that the database then executes, so the result reflects current data rather than model memory. No model training or fine-tuning is involved. Embedding-based similarity search solves a different problem — finding semantically related content — and would not produce an exact aggregate. And Select AI is not an optimizer hint or plain SQL syntax extension; it is explicitly an LLM-backed natural-language interface.