OCI GenAI Pro key terms

A neural network trained on large text corpora to predict tokens and generate language. In OCI it is accessed through the Generative AI service rather than hosted by the customer.

A subword unit of text produced by a model's tokenizer. Billing, context limits, and latency are all measured in tokens, so prompt length directly affects cost and throughput.

The maximum number of tokens a model can consider in a single request, covering both the prompt and the generated output. Exceeding it causes truncation or an error.

Designing instructions, examples, and formatting in the input text to steer model output. It requires no training and is the fastest way to improve response quality.

Supplying a small number of worked input-output examples inside the prompt so the model imitates the desired pattern. Useful when fine-tuning is unnecessary or too costly.

A leading instruction that sets persona, tone, and behavioral constraints for a conversation, applied before user turns. It is a soft control and must not be relied on as a security boundary.

A sampling parameter that scales randomness. Lower values make output more deterministic and factual; higher values increase diversity. Set near zero for extraction and classification tasks.

Nucleus sampling that restricts generation to the smallest set of tokens whose cumulative probability reaches P. It is an alternative diversity control to temperature, usually tuned with it, not independently.

A request parameter capping how many tokens the model may generate. It bounds latency and cost but can truncate answers, so size it against the expected output length.

A general-purpose model trained on broad data before any task-specific adaptation. OCI Generative AI offers several families in different sizes so you can trade quality against cost and latency.

Choosing among available model families and parameter sizes to balance answer quality, latency, context length, and price per token. Larger models generally cost more and respond more slowly.

On-demand inference bills per token with shared capacity and no commitment. Dedicated AI clusters reserve compute for predictable throughput, fine-tuning, or hosting custom models, billed by time.

Further training a pretrained model on your own labeled examples so it learns a task, style, or domain vocabulary. It changes model weights and requires a dedicated cluster to host the result.

A parameter-efficient tuning method offered by OCI Generative AI that updates a small fraction of weights. It reduces training cost and data requirements compared with full fine-tuning.

The curated prompt-and-completion records used for fine-tuning, stored in Object Storage. Quality, coverage, and consistent formatting matter more than raw volume for good results.

A confident model statement that is unsupported or factually wrong. It arises because models predict plausible text, not verified truth, and is mitigated by grounding, citations, and evaluation.

A dense numeric vector representing text meaning, produced by an embedding model. Similar texts land close together, enabling semantic search and clustering without keyword matching.

A model that converts text into vectors for similarity comparison. The same model and version must embed both documents and queries, or distances become meaningless.

A store that indexes embeddings and answers nearest-neighbor queries efficiently. OCI offers a vector store capability alongside database services that support vector columns and indexes.

An indexing strategy that returns close-but-not-guaranteed nearest vectors to gain speed at scale. Recall versus latency is a tuning tradeoff governed by index parameters.

Retrieval ranked by meaning similarity in embedding space rather than exact keyword overlap. It finds paraphrases and synonyms but can miss rare identifiers better served by lexical search.

Combining lexical keyword scoring with vector similarity, then merging results. This improves recall for exact terms such as part numbers while retaining semantic matching.

Splitting source documents into passages small enough to embed and fit in the prompt. Chunk size and overlap strongly affect retrieval precision and answer completeness.

A pattern where relevant passages are retrieved and inserted into the prompt so the model answers from supplied evidence. It reduces hallucination and keeps answers current without retraining.

Anchoring generated answers in retrieved or supplied source content, ideally with citations. It lets reviewers verify claims and limits unsupported invention.

A second-stage scoring of initially retrieved passages with a more precise model before prompting. It improves the order and relevance of the few chunks that enter the context window.

Controls that constrain model input and output, such as content filters, blocked topics, or required formats. They reduce harmful or off-policy responses but are not a complete security solution.

An attack where untrusted text, such as a retrieved document, instructs the model to ignore its rules or leak data. Mitigate with least privilege, input isolation, and output validation.

Detecting and filtering harmful or disallowed content in prompts and responses. It is typically one layer within a broader responsible-AI and safety design.

Design principles covering fairness, transparency, privacy, safety, and human oversight for AI systems. In practice it means documenting intended use, limitations, and evaluation evidence.

Measuring output quality against defined criteria using test sets, human review, or automated scoring. It should cover accuracy, groundedness, safety, and latency before and after deployment.

OCI's authorization system of users, groups, policies, and compartments. Access to Generative AI resources, endpoints, and data is granted through policies scoped to compartments.

A logical container for OCI resources used to organize and isolate them for quota, cost tracking, and access control. Policies written at one compartment can be inherited by children.

Granting only the permissions needed for a task, and no more. For AI workloads this means scoping model invocation, data access, and cluster management to separate, minimal policies.

A network access path that keeps traffic to a service inside a virtual cloud network instead of the public internet. It reduces exposure for model invocation and data retrieval.

An OCI service for storing and controlling encryption keys and secrets. It supports encrypting stored data and managing credentials used by AI applications.

Collecting logs, metrics, and traces to understand system behavior in production. For generative AI it should capture latency, token usage, errors, and retrieval quality signals.

Managing spend through quotas, per-team compartments, token budgets, caching, and model right-sizing. Without limits, long prompts and retries can silently multiply inference charges.