Generative AI Fundamentals and the OCI Generative AI Service
Domain 1 - Generative AI Fundamentals and the OCI Generative AI Service
This domain tests whether you can explain how large language models (LLMs) behave and map that behavior onto the OCI Generative AI service's building blocks: the playground, on-demand inference, dedicated AI clusters, endpoints, and the model catalog. Expect terminology questions (tokens, context window, embeddings, hallucination) plus service-shape questions about what runs on shared versus dedicated infrastructure.
Core rules
- An LLM predicts the next token; the context window bounds prompt plus completion. Cost and latency scale with tokens.
- Sampling controls: temperature scales randomness; top_p/top_k truncate the candidate distribution. Use temperature near 0 for classification or extraction, and do not tune temperature and top_p aggressively at the same time.
- Base models are served on-demand on shared capacity. Fine-tuned or custom-hosted models require a dedicated AI cluster; an endpoint binds a model to capacity.
- OCI hosts Cohere Command R/R+ and Embed, Meta Llama, and Mistral families. Model availability is regional, so data-residency requirements can eliminate a model.
- OCI states that customer prompts and completions are not used to train the underlying models.
Worked example
A team must classify 40,000 support emails nightly into 12 categories with stable output. Reasoning: classification needs determinism, so set temperature 0 and a small max_tokens, and constrain the label set in the prompt. Volume is batch, so cost favors on-demand inference over a dedicated cluster. Because the task is narrow and labeled data is scarce, first try few-shot prompting; only move to T-Few fine-tuning if accuracy is insufficient, which then requires a dedicated fine-tuning cluster.
Traps
Confusing the playground (testing) with an endpoint (production invocation); assuming a fine-tuned model can be called on-demand; memorizing model names without their regional availability; treating temperature as a quality knob rather than a randomness knob.