1Z0-1127-25 Practice Exam 55 — 20 Questions

Question 1

A developer is building a LangChain agent that must maintain memory across multiple turns in a conversation. The agent uses a custom tool to query an Oracle Database 23ai table. After several interactions, the agent starts repeating questions it already asked. Which component in the LangChain agent architecture is most likely responsible for retaining conversational context?

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

Accepted Answer

The agent's memory module. The memory module in LangChain is responsible for persisting and retrieving conversational context. When an agent repeats questions, it indicates that prior turns are not being stored or accessed. The prompt template, tools, and output parser are stateless components that do not retain dialogue history. Therefore, the memory module is the correct component to address the issue.

Answer

The agent's prompt template

Answer

The agent's output parser

Answer

The agent's tool definitions

Question 2

A team is building a document Q&A application on OCI Generative AI. They want the model to answer using only retrieved passages, cite the passage identifier after each claim, and say "insufficient context" when retrieval returns nothing relevant. Which prompt structure BEST supports all three requirements?

Accepted Answer

A prompt with sections for retrieved passages (each labeled with an ID), an instruction to answer only from those passages, a citation format rule, and a fallback phrase for empty retrieval.. A grounded RAG prompt needs labeled evidence, an explicit scope restriction, a defined citation format, and a fallback for missing context. Together these enable passage-only answers with traceable citations and a safe response when retrieval returns nothing relevant. Generic instructions, chain-of-thought without evidence, and knowledge blending all fail at least one requirement.

Answer

A prompt that lists retrieved passages but instructs the model to blend them with its own general knowledge for a richer answer.

Answer

A single instruction to "answer the question accurately" followed by the raw user question.

Answer

A prompt that instructs the model to "think step by step" and then answer the question without providing any retrieved text.

Question 3

A team is evaluating whether to use a pretrained foundation model as-is or to fine-tune it for an internal support assistant. They have only a few hundred high-quality examples and need results quickly. What should they do?

Accepted Answer

Use the pretrained model directly with well-crafted prompts and few-shot examples. A small, high-quality example set plus a fast turnaround points to prompt engineering with few-shot examples on a pretrained model. That path delivers immediate, measurable quality without a training cycle. Full retraining, from-scratch training, and cluster provisioning all demand more data, time, or budget than the situation allows and do not directly improve task accuracy.

Answer

Fine-tune using a full-parameter training job on a dedicated AI cluster

Answer

Deploy a dedicated AI cluster and rely on its isolation to improve response quality

Answer

Train a custom model from scratch using the internal support transcripts as the corpus

Question 4

A solutions architect is configuring a generation request against an OCI Generative AI Cohere Command model. The team wants short, deterministic answers for a FAQ bot where the same question should always produce the same response. Which parameter change BEST supports this goal?

Accepted Answer

Set temperature to 0 and lower top_p to a small value such as 0.1.. Deterministic, repeatable generation is achieved by suppressing sampling randomness. Temperature 0 forces greedy decoding, and a small top_p narrows the candidate set further. Penalties, length caps, and higher temperatures all either preserve or increase variability, so they work against the FAQ bot's requirement.

Answer

Increase max_tokens so the model has more room to settle on a stable answer.

Answer

Set the frequency penalty to a high value so repeated tokens are discouraged.

Answer

Increase temperature to 1.0 and keep top_p at the default.

Question 5

A prompt engineer is tuning a Cohere Command model on OCI Generative AI for a brainstorming tool. The tool must produce diverse marketing slogans while remaining on-brand and never inventing product features. Which TWO prompt engineering practices BEST balance creativity with factual constraint? (Choose two.)

Accepted Answer

Raise the temperature moderately and increase top_p to widen the sampling pool for more varied phrasing.. Balancing creativity and factual constraint requires two complementary moves: an explicit evidence boundary that limits claims to approved features, and a sampling configuration that widens stylistic variety. Constraining content while loosening token selection lets the model explore phrasing without inventing capabilities. Deterministic sampling suppresses diversity, and stop sequences or low top_k neither enforce facts nor support rich brainstorming.

Answer

Use a very low top_k so only the single most likely token is chosen at each step.

Answer

Set temperature to 0 and rely on the model's default product knowledge for factual accuracy.

Answer

Add a stop sequence for the word "feature" so the model cannot discuss product capabilities.

Question 6

A developer is calling the OCI Generative AI chat endpoint with a Cohere Command R model and wants the model to produce more deterministic, repeatable answers for a structured data extraction task. They set temperature to 0. Which additional parameter should they also tune to reduce randomness in token selection?

Accepted Answer

Top-p (nucleus sampling). Temperature and top-p are the two primary sampling controls. With temperature at 0, lowering top-p narrows the candidate token set to the most probable tokens, which reduces randomness and produces more consistent, repeatable extraction results.

Answer

Max tokens

Answer

Presence penalty

Answer

Frequency penalty

Question 7

A financial services team uses an OCI Generative AI Cohere Command model to summarize quarterly earnings call transcripts. They want summaries that stay strictly grounded in the transcript and refuse to speculate on future revenue. Which prompt engineering technique should they apply to BEST achieve this?

Accepted Answer

Use a system prompt that instructs the model to answer only from the provided transcript and to respond "not stated" if information is absent.. Grounding a model in a specific source document is achieved through explicit prompt instructions that define the allowed evidence and a fallback behavior when the answer is missing. A system prompt scoping the model to the transcript and requiring a "not stated" response prevents speculation. Sampling parameters like temperature, top_k, and max_tokens affect randomness and length, not factual grounding.

Answer

Lower the top_k value to 1 so the model always picks the single most probable token.

Answer

Set the temperature parameter to 1.0 and add "be creative" to the prompt.

Answer

Increase the max_tokens parameter to the model's maximum so the summary can include every detail.

Question 8

A RAG application on OCI Generative AI returns answers that cite irrelevant or outdated chunks. You have verified that the embedding model is appropriate and the vector index is healthy. Which TWO techniques should you apply to improve the relevance of retrieved context? (Choose two.)

Accepted Answer

Add metadata filters to restrict retrieval to documents matching the user's department and recency.. Metadata filtering and reranking both improve the precision of retrieved context. Filtering removes candidates that are irrelevant by attributes such as department or date, while reranking reorders the top candidates using a more accurate scoring model. Together they ensure the LLM receives focused, current, and relevant chunks, reducing citations of irrelevant or outdated material.

Answer

Re-embed the entire corpus using the same embedding model to refresh the vectors.

Answer

Increase the LLM temperature to encourage more diverse answers.

Answer

Reduce the number of chunks sent to the LLM to one per query.

Question 9

A data scientist is preparing a prompt for an OCI Generative AI model to classify support tickets into categories. The model performs well on the first few tickets but then drifts, mixing up labels. The scientist wants to improve reliability by giving the model clear examples of each category within the prompt itself, without retraining any weights. Which technique should be applied?

Accepted Answer

Few-shot prompting, by including several labeled input-output examples for each category in the prompt. Providing a few labeled examples inside the prompt is in-context learning, specifically few-shot prompting. The demonstrations let the model induce the classification mapping and output format at inference time with no gradient updates, which stabilizes labels and satisfies the no-retraining requirement. Fine-tuning, retrieval, and decoding changes address different problems.

Answer

Beam search decoding with a wider beam to explore more candidate label sequences

Answer

Retrieval-augmented generation, by embedding all historical tickets and retrieving similar ones at inference

Answer

Fine-tuning the model on a labeled ticket dataset to bake the categories into the weights

Question 10

An engineer notices that an LLM occasionally generates answers that are fluent and confident but contradict the provided source document. The team wants to reduce this behavior without retraining the model. Which technique is most appropriate?

Accepted Answer

Lower the temperature and instruct the model to answer only from the provided document. Lowering temperature reduces random sampling that can pull the model away from the source, and an explicit instruction to answer only from the provided document constrains generation to the given context. These prompt-level changes reduce contradictions without retraining, unlike raising temperature, changing model size, or removing the source.

Answer

Remove the source document from the prompt to avoid confusing the model

Answer

Add more parameters to the model by switching to a larger variant

Answer

Increase the temperature to encourage more diverse outputs

Question 11

A machine learning engineer is fine-tuning a large language model on OCI Generative AI. The training job completes, but the resulting model produces repetitive and nonsensical outputs when given new prompts. The engineer suspects overfitting. Which training hyperparameter adjustment is most likely to mitigate this issue?

Accepted Answer

Decrease the number of training epochs. Overfitting in fine-tuning often manifests as repetitive or nonsensical outputs on new data. Reducing the number of training epochs is a direct way to prevent the model from memorizing the training set. This allows the model to capture general patterns rather than noise, improving generalization to unseen prompts.

Answer

Increase the model's context window

Answer

Increase the learning rate

Answer

Increase the batch size

Question 12

A developer wants to embed OCI Generative AI chat completions directly into a Python application that already uses the OCI SDK for other services, without adding a new SDK family. Which approach fits this requirement?

Accepted Answer

Use the OCI Python SDK's Generative AI client with the chat API and the configured OCI config profile. Reusing the OCI Python SDK's Generative AI client keeps authentication, retries, and configuration consistent with the rest of the application and exposes chat completions without adding another SDK. Alternatives either target a different platform's API, add process overhead, or use an authentication model OCI does not accept for these endpoints.

Answer

Call the model's REST endpoint with a bearer token issued by an external identity provider

Answer

Install the Cohere Python SDK and call the Cohere endpoint with an OCI API key

Answer

Use the OCI CLI to invoke chat completions and parse the JSON output inside the Python process

Question 13

A developer is building a LangChain pipeline against OCI Generative AI. They need a component that accepts a natural-language question plus a set of retrieved Documents and returns a single string answer, so it can be dropped directly into an LCEL pipeline with a RunnablePassthrough for the question and a retriever for the documents. Which LangChain runnable should they instantiate for this role?

Accepted Answer

create_stuff_documents_chain. The requirement is a runnable whose inputs are retrieved Documents plus a question and whose output is a plain string answer. The stuff-documents chain constructor produces exactly that: it stuffs the document contents into a prompt, calls the model, and returns text. The retrieval chain and history-aware retriever operate at different layers, and the passthrough utility only manipulates dictionary keys.

Answer

RunnablePassthrough.assign

Answer

create_history_aware_retriever

Answer

create_retrieval_chain

Question 14

A developer is configuring an OCI Generative AI text-generation call and wants the model to produce a deterministic, repeatable answer for the same prompt every time, with no creative variation. Which parameter should be set to zero?

Accepted Answer

Temperature. Temperature divides the logits before softmax; at zero the highest-scoring token receives essentially all the probability, producing greedy decoding. Identical prompts then yield identical completions, which is exactly the repeatable behavior requested. Other sampling controls only reshape which tokens are eligible and do not eliminate randomness on their own.

Answer

Max tokens

Answer

Frequency penalty

Answer

Top-p

Question 15

A platform team is evaluating an OCI Generative AI deployment for a customer-facing assistant. They must reduce the risk of the model producing fabricated facts and of leaking sensitive prompt content. Which two techniques should the team implement to address these risks? (Choose two.)

Accepted Answer

Ground responses in retrieved enterprise documents and instruct the model to answer only from the supplied context, declining when the context is insufficient.. Grounding generation in retrieved documents with an explicit abstain instruction constrains the model to verifiable source material, which curbs fabricated facts. Guardrails that inspect and redact sensitive entities on both the request and the response prevent confidential values from reaching the model or the user, covering the leakage risk. Together they address both concerns.

Answer

Increase the context window by selecting a model with the largest available token limit.

Answer

Raise the temperature and top_p values so the assistant produces more varied and natural-sounding answers.

Answer

Disable logging on the inference endpoint so prompts and completions are never stored.

Question 16

A team is building an LCEL RAG pipeline over OCI Generative AI. They want to stream partial tokens to the browser as the model generates the final answer, and they also want to keep the retrieved source Documents available for citation display. Which TWO design choices support these goals? (Choose two.)

Accepted Answer

Use create_retrieval_chain so the final result includes both the generated answer and the retrieved source documents.. Streaming requires calling the streaming interface on the runnable that wraps the chat model, so token fragments can be forwarded as generated. Citations require a composition whose output retains the retrieved Documents next to the answer. Combining a streaming answer runnable with a retrieval chain that returns sources satisfies both, while blocking invocation, retriever-level streaming, and parallel wrappers do not.

Answer

Call invoke() on the answer chain and poll the returned object for partial tokens while it is still executing.

Answer

Enable streaming on the vector store retriever so document chunks are emitted token by token during similarity search.

Answer

Wrap the entire LCEL chain in RunnableParallel so every branch shares one streaming connection to the model.

Question 17

A team's RAG application on OCI Generative AI is returning correct but incomplete answers: users report that facts present in the uploaded documents are being missed. The vector store ingestion completed successfully and the model endpoint is healthy. Which tuning step most directly addresses missed relevant passages at query time?

Accepted Answer

Increase the number of retrieved chunks passed to the model and verify the chunk size and overlap used during ingestion.. Incomplete answers that omit facts known to be in the documents point at retrieval rather than generation, so the effective levers are the number of retrieved chunks and how documents were chunked during ingestion. Output limits, deployment mode, and authorization scope do not change which passages reach the model.

Answer

Apply a stricter IAM policy that limits which documents the application can read at inference time.

Answer

Lower the model's maximum output tokens so answers stay focused on the retrieved context.

Answer

Switch the endpoint from a dedicated AI cluster to on-demand capacity to reduce retrieval latency.

Question 18

A team is evaluating an LLM for a retrieval-augmented generation (RAG) application on OCI. They observe that when the retrieved context is long, the model sometimes ignores information placed in the middle of the context window, even though it answers correctly when the same information appears at the beginning or end. Which phenomenon best explains this behavior?

Accepted Answer

The 'lost in the middle' effect due to positional bias in attention.. The described behavior matches the well-documented 'lost in the middle' effect, where LLMs exhibit a U-shaped recall curve across long contexts. Information at the start and end of the input receives stronger attention than content buried in the middle. For RAG on OCI, this means retrieved chunks should be ordered so the most relevant passages appear near the beginning or end of the prompt, improving answer fidelity without changing the model or retriever.

Answer

Catastrophic forgetting during pretraining.

Answer

Insufficient embedding dimensionality in the vector store.

Answer

Tokenization mismatch between the query and the retrieved documents.

Question 19

A developer is building a LangChain agent that must maintain memory across multiple turns in a conversation. The agent uses a custom tool to query an Oracle Database 23ai table. After several interactions, the agent starts repeating questions it already asked. Which component in the LangChain agent architecture is most likely responsible for retaining conversational context?

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

Accepted Answer

The AgentExecutor, which runs the agent loop, invokes tools, and feeds observations back to the model until it produces a final answer.. The AgentExecutor is the runtime that implements the reason-act-observe cycle: it calls the agent, executes the chosen tool, appends the observation to the scratchpad, and repeats until a final answer is produced. Its configuration options, including iteration limits and parsing-error handling, directly address the requirement to recover from tool failures and keep reasoning.

Answer

The Memory module, which stores intermediate tool outputs and replays them to the model on each iteration.

Answer

The Retriever, which decides which tool to invoke by performing a similarity search over the agent's tool descriptions.

Answer

The OutputParser, which converts the model's text into a structured AgentAction or AgentFinish object that terminates the loop.

Question 20

A company's OCI Generative AI fine-tuning job on a Cohere Command R model completed successfully, but evaluation shows the model performs poorly on the target task. The dataset contains 500 high-quality examples, and the team used the default hyperparameters. Which action is MOST likely to improve results?

Accepted Answer

Increase the number of training epochs and add more domain-specific examples. Poor performance after fine-tuning with a small dataset and default hyperparameters usually indicates insufficient training signal. Increasing epochs and adding more high-quality, domain-specific examples directly improves the model's ability to learn the task and is the most likely fix.

Answer

Change the base model to a smaller model to reduce inference cost

Answer

Reduce the dataset to 100 examples to avoid overfitting

Answer

Switch from fine-tuning to prompt engineering only, and discard the fine-tuned model