Overview
Domain 1 covers the model behaviour that every later section assumes. You need to explain how a decoder-only transformer predicts the next token, why a context window caps prompt plus output, and how tokens differ from words. The section then moves to the settings you pass on each call, such as temperature, top-p, top-k, maximum output tokens and frequency or presence penalties, and to prompt techniques like few-shot examples, system instructions and step-by-step reasoning. Embeddings, semantic similarity and retrieval-augmented generation close the loop with grounding.
Exam status: Oracle's certification index lists Oracle Cloud Infrastructure Enterprise AI Professional (1Z0-1158-26) and no longer lists OCI Generative AI Professional (1Z0-1127). Our catalog treats 1Z0-1158-26 as the successor, but we found no Oracle statement that formally announces the replacement, so confirm the status on Oracle University before you book.
How it's tested
Language model questions tend to be short scenarios: a team sees repetitive output, invented facts or truncated replies, and you pick the parameter, prompt change or architecture that fixes it. Some ask you to define a term, but most test whether you can connect a symptom to its cause.
-
Symptom to setting
Output is too random, too repetitive or cut short, and you choose which inference parameter to change and in which direction.
-
Grounding choice
A model invents policy details, and you decide whether retrieval, a better prompt or fine-tuning addresses the gap.
-
Prompt technique
You match a goal such as consistent format or multi-step arithmetic to few-shot examples, a system instruction or reasoning prompts.
A support bot gives a different refund policy on each run and sometimes invents a 90-day window. Two problems are mixed here. Variation between runs points to sampling, so lower the temperature. Invented policy points to missing knowledge, so ground the model with retrieved policy text. Fine-tuning would not keep facts current, and raising top-k would make variation worse.
Topics 5
Official topics from the Oracle outline, in outline order.
1.1 Explain how large language models generate textTransformers, tokens, next-token prediction and context windows.
- What it means
- A large language model is a transformer network trained to predict the next token given the tokens before it. Text is split into tokens, which are often word fragments, and the model produces one token at a time until it reaches a stop condition or a length limit.
- What to understand
- The context window is the total token budget for input plus generated output in one call. Long histories, retrieved passages and tool results all consume it. When the budget is exceeded, requests fail or older content must be trimmed or summarised. Token counts also drive on-demand cost, so they link this topic to optimization questions later in the exam.
- How it shows up
- Expect a question about why a long multi-turn exchange loses early instructions or why a reply stops mid-sentence.
- Common confusion
- Treating tokens as whole words, which leads to wrong estimates of context use and cost.
1.2 Control output with inference parametersTemperature, top-p, top-k, max tokens and penalties.
- What it means
- Inference parameters change how the next token is chosen. Temperature reshapes the probability distribution, top-k limits sampling to the k most likely tokens, top-p limits it to the smallest set whose probabilities add up to p, and maximum output tokens caps reply length.
- What to understand
- Low temperature gives focused, repeatable answers suited to extraction and classification, while higher values give variety for drafting. Frequency penalty discourages tokens in proportion to how often they have already appeared; presence penalty discourages any token that has appeared at all. A low max tokens value truncates output and does not make the model more concise in style. Know which parameters each supported model accepts, because not every model exposes all of them.
- How it shows up
- A scenario describes repetitive, random or truncated output and asks which setting to adjust.
- Common confusion
- Mixing up top-p and top-k, or thinking temperature zero guarantees identical output on every model.
1.3 Apply prompt engineering techniquesSystem instructions, zero-shot, few-shot and reasoning prompts.
- What it means
- Prompt engineering shapes behaviour without changing model weights. A system or developer instruction sets role and rules, zero-shot prompting gives only the task, few-shot prompting adds worked examples, and reasoning prompts ask the model to work through intermediate steps before answering.
- What to understand
- Few-shot examples are the cheapest way to fix output format and labelling style. Clear delimiters separate instructions from untrusted content, which also reduces prompt injection risk. Reasoning-capable models may plan internally, so you choose them rather than forcing long chain-of-thought text. Asking for structured output, such as JSON that follows a schema, makes downstream parsing reliable. Every example you add costs tokens on every call.
- How it shows up
- You may need to pick the lightest change that makes a model return a consistent format.
- Common confusion
- Reaching for fine-tuning when a few examples in the prompt would solve a formatting problem.
1.4 Use embeddings and retrieval-augmented generationSemantic vectors, similarity search and grounded answers.
- What it means
- An embedding model converts text into a vector so that passages with similar meaning sit close together. Retrieval-augmented generation embeds a question, finds the nearest stored chunks, and places them in the prompt so the model answers from that evidence.
- What to understand
- RAG keeps answers current without retraining, since you update documents rather than weights. Quality depends on chunk size, the embedding model, the number of passages returned and whether the prompt tells the model to rely on them. Query and documents must be embedded with one model. Citations let users check sources. Hybrid retrieval that adds keyword matching helps with product codes and exact names that pure semantic search can miss.
- How it shows up
- Questions ask which approach answers from frequently changing internal documents with source references.
- Common confusion
- Thinking RAG changes the model; it only changes what the model sees at inference time.
1.5 Choose between prompting, retrieval and fine-tuningMatching the customization method to the problem.
- What it means
- There are three main levers for adapting a foundation model: better prompts, retrieval of external knowledge, and fine-tuning on labelled examples. Each solves a different kind of gap and has different cost, latency and maintenance profiles.
- What to understand
- Prompting is fastest and cheapest to try. Retrieval fits knowledge gaps and changing facts. Fine-tuning fits gaps in style, tone, format or a narrow task where many examples exist and prompts have reached their limit. It needs training data, compute and a hosting endpoint, and it bakes knowledge in at training time. Many production designs combine a tuned model with retrieval. Exam answers usually favour the least effort option that meets the stated requirement.
- How it shows up
- A scenario lists data volume, freshness and style needs and asks which lever to pull first.
- Common confusion
- Assuming fine-tuning is the right way to teach a model new facts that change monthly.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Token
The unit a model reads and writes, often part of a word.
- Remember
- Context limits and on-demand pricing are counted in tokens.
- Tested as
- Cost and truncation questions.
- Trap
- Estimating usage by word count.
-
Context window
Maximum tokens for prompt plus output in one request.
- Remember
- History, retrieved text and tool results all use it.
- Tested as
- Long conversation and RAG scenarios.
- Trap
- Forgetting output tokens count too.
-
Temperature
Scales how sharply the model favours likely tokens.
- Remember
- Lower means more deterministic; higher means more varied.
- Tested as
- Consistency questions.
- Trap
- Raising it to fix factual errors.
-
Top-p
Samples from the smallest token set reaching cumulative probability p.
- Remember
- Adapts the candidate count to the distribution.
- Tested as
- Sampling comparison items.
- Trap
- Confusing it with a fixed token count.
-
Top-k
Samples only from the k most probable tokens.
- Remember
- Fixed size regardless of probabilities.
- Tested as
- Sampling comparison items.
- Trap
- Thinking larger k makes output safer.
-
Frequency and presence penalties
Reduce repetition of tokens already generated.
- Remember
- Frequency scales with count; presence applies once.
- Tested as
- Repetitive output scenarios.
- Trap
- Treating them as the same setting.
-
Few-shot prompting
Including worked input-output examples in the prompt.
- Remember
- Cheapest fix for format and labelling consistency.
- Tested as
- Prompt design questions.
- Trap
- Ignoring the per-call token cost.
-
Hallucination
Fluent output that is not supported by facts.
- Remember
- Grounding with retrieved sources is the main mitigation.
- Tested as
- Accuracy and trust scenarios.
- Trap
- Expecting low temperature alone to remove it.
-
Embedding
A vector that captures the meaning of text.
- Remember
- Similar meaning yields nearby vectors.
- Tested as
- RAG and semantic search questions.
- Trap
- Mixing vectors from two models.
-
Retrieval-augmented generation
Fetching relevant passages and adding them to the prompt.
- Remember
- Keeps answers current without retraining.
- Tested as
- Knowledge freshness scenarios.
- Trap
- Believing it updates model weights.
-
Fine-tuning
Further training a base model on labelled examples.
- Remember
- Best for style, format or narrow task behaviour.
- Tested as
- Customization choice questions.
- Trap
- Using it to inject frequently changing facts.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Sampling parameters
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Temperature | Sharpens or flattens the token distribution | Applies to every token choice | Thinking zero is perfectly deterministic everywhere |
| Top-k | Keeps a fixed number of candidates | Count-based cutoff | Swapping its meaning with top-p |
| Top-p | Keeps candidates up to a probability mass | Probability-based cutoff | Setting it to 1 and expecting focus |
| Max output tokens | Upper bound on generated length | Truncates rather than condenses | Using it to request shorter style |
Ways to adapt a model
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Prompt engineering | Instructions and examples in the request | No training, instant iteration | Overloading prompts until context runs out |
| Retrieval-augmented generation | Inject retrieved passages at run time | Fresh, citable knowledge | Poor chunking blamed on the model |
| Fine-tuning | Adjust weights with labelled pairs | Changes style and task behaviour | Expecting facts to stay current |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Temperature fixes accuracy
- What candidates pick
- Lower temperature to stop invented policy details.
- Why it looks right
- It makes answers more stable.
- Why it's wrong
- Stable answers can still be wrong; missing knowledge is solved by grounding with retrieved sources, not by sampling changes.
Clue to watch for The scenario says facts are absent from training data or recently changed.
-
Max tokens makes answers concise
- What candidates pick
- Cut max output tokens so replies become shorter summaries.
- Why it looks right
- Fewer tokens sounds like brevity.
- Why it's wrong
- The limit simply stops generation, often mid-sentence; brevity comes from instructions or examples.
Clue to watch for Users complain about replies ending abruptly.
-
Fine-tune for fresh facts
- What candidates pick
- Fine-tune monthly on new product data.
- Why it looks right
- Training feels like the thorough option.
- Why it's wrong
- Retraining is slow and costly, and tuned facts go stale; retrieval keeps content current at lower effort.
Clue to watch for Words such as frequently updated, latest or catalogue changes.
-
Bigger context is always better
- What candidates pick
- Stuff every document into the prompt.
- Why it looks right
- More evidence seems safer.
- Why it's wrong
- Extra text raises cost and latency and can bury the relevant passage; targeted retrieval of top chunks works better.
Clue to watch for The prompt mentions hundreds of pages or rising token bills.
-
Penalties are interchangeable
- What candidates pick
- Use presence penalty to scale with repeat count.
- Why it looks right
- Both reduce repetition.
- Why it's wrong
- Only frequency penalty grows with how often a token appears; presence penalty is a flat one-time adjustment.
Clue to watch for The question stresses how many times a phrase repeats.
Answer framework
Work through these steps when a model output problem appears in a 1Z0-1158-26 scenario.
- Classify the symptom as randomness, repetition, truncation, wrong format or wrong facts, because each has a different family of fixes.
- For randomness or repetition, adjust temperature, top-p, top-k or penalties before touching architecture or training.
- For format problems, add a clear instruction, a schema request or a few examples to the prompt and measure again, keeping the instruction and examples consistent so you can tell which change actually improved the output.
- For missing or stale knowledge, add retrieval over trusted sources and instruct the model to answer only from them.
- Consider fine-tuning only when many labelled examples exist and the gap is style or task behaviour that prompts cannot close; a fine-tuned model must then be hosted on its own endpoint before callers see the new behaviour.
Practice question
Want more like this? The 1Z0-1158-26 practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Begin with the playground in the OCI Console, or a short script against an on-demand chat model, and run the same prompt ten times at temperature 0.1 and then at 0.9. Record what changes. Repeat with top-p and top-k so the difference between a probability cutoff and a count cutoff becomes concrete. Set a tiny max output token value and watch the reply stop mid-sentence.
Next, write one classification prompt three ways: zero-shot, few-shot with three examples, and with a JSON schema request. Compare consistency and token counts. Then build a minimal retrieval demo with a handful of documents, so you can see grounding fix an invented answer that temperature alone could not.
Finish by writing a one-page table of symptoms and fixes: randomness, repetition, truncation, format drift, stale facts. Add a column for whether prompting, retrieval or fine-tuning is the lightest fix. Review it before attempting practice questions, because most items in this section test that mapping rather than definitions. Keep this domain to roughly a week of study time and move on to agents.
FAQ
How much of 1Z0-1158-26 is Domain 1, Applying Large Language Model Concepts?
Oracle does not publish section weights. Third-party practice banks give this section one of the smaller shares, roughly a tenth of items, but its terms appear throughout the agent and model sections.
Is this the same content as the old OCI Generative AI Professional exam?
The LLM fundamentals overlap heavily with 1Z0-1127, but 1Z0-1158-26 places them inside agent and Responses API scenarios. Oracle has not published a mapping between the two exams.
Do I need to know transformer mathematics?
No detailed math is expected. Know next-token prediction, tokens, attention at a conceptual level and the context window limit.
Does temperature zero give identical answers every time?
It makes output close to deterministic, but some models and serving stacks still show small differences. Treat it as highly consistent rather than exact.
When should I pick RAG over fine-tuning?
Pick retrieval when knowledge changes or answers need citations. Pick fine-tuning when you need a consistent style or narrow task behaviour backed by many labelled examples.
Are embeddings tested in this section?
Yes at a concept level: what a vector represents, similarity search and the rule that queries and documents share one embedding model.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
You have covered tokens, context windows, sampling parameters, prompt techniques, embeddings, retrieval and customization choices. See how well you map symptoms to fixes.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the 1Z0-1158-26 AI tutor Get any concept explained your way