OCI Generative AI Professional (1Z0-1127) - Practice Questions¶
15 questions for OCI Generative AI Professional prep. Three decisions recur: how to customize, on-demand versus dedicated cluster, and where retrieval quality comes from.
Cert page: exams/oracle/oci-generative-ai-professional/
Question 1¶
Scenario: An insurer wants an assistant answering questions about policy documents. Documents change monthly, answers must cite the source clause, and there are no ML engineers.
A. Fine-tune a base model on the policy corpus each month so the clause wording is learned into its weights B. Pre-train a domain model on the insurer's own documents from scratch C. RAG over the policy documents, citing the retrieved clause D. Put all documents in the prompt on every request and rely on the model to find the clause
Answer
**Correct: C** **Why:** Monthly updates rule out fine-tuning, whose knowledge is frozen at training time. Citation requires knowing which source text produced the answer, which only retrieval provides. Putting all documents in the prompt exceeds the context window and costs enormously.Question 2¶
Scenario: A team has fine-tuned a model and wants to serve it to production with a latency commitment.
A. Either, depending on volume B. A dedicated AI cluster, which is required to host a custom model and gives reserved capacity for latency C. OCI Data Science model deployment, the managed service for serving fine-tuned models D. On-demand inference
Answer
**Correct: B** **Why:** On-demand serves base models only. Both fine-tuning and hosting a custom model require dedicated AI clusters, and dedicated capacity is also what makes a latency commitment possible. This is a hard constraint, not a preference.Question 3¶
Scenario: A RAG chatbot gives confidently wrong answers although the information is in the corpus. The team increased retrieved chunks from 3 to 20 and it got worse.
A. Raise the temperature so the model explores other phrasings and is less anchored to the wrong passages it retrieved B. Log the retrieved chunks to check the right one is present, then cut the count and consider reranking C. Increase the chunk count further, to 40, so that more of the corpus reaches the model and the right passage is certain to be included D. Fine-tune the model on the corpus
Answer
**Correct: B** **Why:** Diagnose before changing. If the correct chunk is retrieved but ignored, 20 chunks is likely the cause because models attend less to the middle of a long context. If it is not retrieved, the problem is chunking, the embedding model, or query phrasing. Raising temperature increases variety, not correctness.Question 4¶
Scenario: A pipeline extracts structured fields from contracts and produces slightly different values on repeated runs, breaking downstream reconciliation.
A. Set temperature to 0 and validate the output against a schema B. Fine-tune the model on labelled contracts so that it always returns identical values for the same contract text C. Increase max tokens so that the model has room to finish every extraction D. Retry each extraction until two runs agree, then keep the shared value
Answer
**Correct: A** **Why:** Temperature is the primary control over randomness, and 0 gives near-deterministic decoding. Schema validation is the safety net, since output is not guaranteed identical across model versions. Fine-tuning is an expensive answer to a decoding parameter problem.Question 5¶
Scenario: Inference costs tripled. Every request sends the full conversation history, 15 retrieved chunks, and a 2,000-token system prompt.
A. Reduce retrieved chunks, window or summarize memory, shorten the system prompt, cap max tokens B. Reduce traffic C. Cache whole responses D. Move to a dedicated cluster, because a fixed hourly price always works out cheaper than per-token billing at any volume
Answer
**Correct: A** **Why:** Cost is driven by tokens in and out. The largest levers, in order, are retrieved context size, conversation history handling, and system prompt length. A dedicated cluster gives predictable cost, not necessarily lower cost. Response caching helps only for repeated identical questions.Question 6¶
Scenario: A multi-tenant assistant returns one organization's content to a user from another. What is the correct fix?
A. Filter by tenant in the vector query or use per-tenant namespaces, authorizing on user identity B. Deploy a separate fine-tuned model per tenant, with documents baked into the weights C. Post-filter results after generation D. Instruct the model in the system prompt not to reveal other tenants' data
Answer
**Correct: A** **Why:** This is a retrieval authorization bug. A prompt is not an access control. Post-filtering is too late because the data has already entered the model's context. Separate models per tenant is enormously expensive and does not address the retrieval boundary.Question 7¶
Scenario: Which statement about fine-tuning versus RAG is correct?
A. They are alternatives that cannot be combined, since a tuned model cannot take retrieved context B. RAG requires retraining when documents change C. Fine-tuning is always more accurate D. Fine-tuning changes how the model behaves; RAG changes what it knows on each request, so they can be combined
Answer
**Correct: D** **Why:** This is the decision rule the exam applies throughout. RAG updates instantly when source documents change, which is precisely why it beats fine-tuning for knowledge. The two can be combined: a fine-tuned model answering over retrieved context.Question 8¶
Scenario: Chunks are set to 50 tokens each. Answers cite the right documents but lack coherent explanation.
A. Change the similarity metric from cosine to dot product, since the right documents are found but ranked by the wrong measure B. Lower the temperature C. Increase chunk size, add overlap, and split on structural boundaries D. Retrieve fewer chunks
Answer
**Correct: C** **Why:** Chunks that are too small lack surrounding context, so retrieval finds the right words but the model cannot ground a coherent answer. Larger chunks with overlap, split on paragraphs or sections, resolve it. The similarity metric and temperature address different problems.Question 9¶
Scenario: Which LangChain component holds prior conversation turns and is a major cost driver?
A. Output parser that formats the model's reply into a structured object B. Text splitter that cuts the source documents into chunks C. Memory, the component that buffers earlier turns D. Retriever
Answer
**Correct: C** **Why:** Naive memory resends the whole conversation each turn, so token use grows sharply over a long session. Windowed memory keeps only recent turns and summary memory compresses older ones, both of which control the cost.Question 10¶
Scenario: What must be true for a query embedding to retrieve the right chunks?
A. The query must use exactly the same words as the document, since matching is done on shared vocabulary rather than meaning B. The query must be embedded with the same model that indexed the chunks C. The chunks must be sorted alphabetically by their source document title before upload D. The vector store must use Euclidean distance rather than cosine similarity or dot product
Answer
**Correct: B** **Why:** Embeddings from different models occupy different vector spaces, so similarity between them is meaningless. Semantic search exists precisely so that queries need not share wording with documents, and the metric must match how the embedding model was trained.Question 11¶
Scenario: Which parameter should be adjusted to make output less repetitive over a long generation?
A. Top-k set to 1, so the model always picks its single most likely token and cannot loop over the same phrases B. Max tokens C. Temperature set to zero D. Frequency penalty
Answer
**Correct: D** **Why:** Frequency penalty reduces the likelihood of tokens in proportion to how often they have already appeared, which directly targets repetition. Presence penalty discourages tokens that have appeared at all. Top-k of 1 is greedy decoding, which usually increases repetition.Question 12¶
Scenario: A RAG system must be evaluated. What should be measured?
A. Answer quality only, scored end to end by reviewers, because retrieval problems always show up in the final answer B. Retrieval quality and answer quality, measured separately C. Token throughput D. Model perplexity
Answer
**Correct: B** **Why:** They fail for different reasons and need different fixes. Retrieval is measured by whether the correct source appeared in the top K and how highly it ranked; answer quality by faithfulness, relevance, and correctness. Perplexity evaluates a model, not an application.Question 13¶
Scenario: Which parameter-efficient fine-tuning method does OCI Generative AI offer?
A. Distillation B. Quantization C. T-Few D. Full fine-tuning only
Answer
**Correct: C** **Why:** T-Few updates a fraction of the weights, which makes fine-tuning far cheaper and faster than full fine-tuning and reduces the amount of training data required. Distillation and quantization are model compression techniques, not fine-tuning methods.Question 14¶
Scenario: What is the primary risk of retrieving from a corpus that outside parties can write to?
A. Slower retrieval, because every outside write forces a complete re-index that locks the vector store until it finishes B. Higher embedding costs C. Duplicate chunks D. Indirect prompt injection: planted instructions reach the model via retrieved text
Answer
**Correct: D** **Why:** Anything an outsider can write into the corpus can be retrieved into a future prompt, which makes the index a delivery vehicle for instructions. Controlling write access to the corpus and constraining what the application can do are the mitigations.Question 15¶
Scenario: Which is the correct order of the RAG indexing phase?
A. Chunk the documents, embed the chunks, store vectors with metadata B. Store the documents, generate, then embed C. Embed the query, search the vector store, retrieve the top chunks, and generate an answer from them D. Fine-tune the model, deploy it to a dedicated cluster, and then evaluate it
Answer
**Correct: A** **Why:** Indexing runs once, or on document change: chunk, embed, store. The query phase is the other option listed: embed the query, search, retrieve, augment the prompt, generate. Keeping the two phases distinct is the basis for diagnosing where a RAG system fails.Scoring guide¶
- 13-15 correct (85%+): Ready. The pass mark is 68%.
- 10-12 correct (65-80%): Review the customization decision table and the diagnosis order for RAG failures.
- Below 10: Work the scenarios and build the RAG pipeline project in this repo.