AWS GenAI Developer Pro key terms

A large model pretrained on broad data that can be adapted to many downstream tasks through prompting or fine-tuning, rather than trained for one narrow purpose.

Settings such as temperature, top-p, and maximum tokens that control how a model samples and how long its output runs, without changing the underlying weights.

The practice of designing instructions, context, and examples so a model produces accurate, consistent, and appropriately formatted output for a given task.

Including a small number of worked examples inside the prompt so the model can infer the desired pattern, tone, or output format without any weight updates.

A technique that asks a model to reason through intermediate steps before answering, improving accuracy on multi-step problems at the cost of more output tokens.

The basic unit of text a model reads and generates, roughly a word fragment; token counts drive billing, context limits, and latency rather than character counts.

The maximum number of tokens a model can consider in a single request, covering system prompt, retrieved passages, conversation history, and the generated response together.

Delivering generated tokens incrementally as they are produced so users see output immediately, reducing perceived latency for long generations in interactive applications.

An architecture that retrieves relevant documents from an external store and inserts them into the prompt so the model answers from current, private, or verifiable sources.

A numeric vector representing the meaning of text, positioned so semantically similar passages sit close together and can be compared by distance or similarity score.

A database or index built to hold embeddings and answer nearest-neighbor queries quickly, often with metadata filters to restrict results by tenant, date, or category.

Splitting source documents into smaller passages before embedding, balancing chunk size so retrieved context stays focused while preserving enough surrounding meaning to answer questions.

Retrieval that ranks passages by embedding similarity to the query rather than exact keyword overlap, capturing paraphrases and synonyms that literal matching would miss.

Combining keyword-based ranking with vector similarity and merging the scores, which improves recall for exact identifiers, product codes, and rare terms that embeddings handle poorly.

A second retrieval stage that scores candidate passages with a stronger model and reorders them, improving precision before the top results are placed into the prompt.

The degree to which a generated answer is actually supported by the retrieved source passages, as opposed to plausible-sounding content the model produced from its own parameters.

Confidently stated output that is factually wrong or unsupported by any provided source, often arising when retrieval returns weak passages or the prompt invites speculation.

Attaching source identifiers to generated statements so reviewers and automated checks can trace each claim back to a specific retrieved passage.

Further training a pretrained model on task-specific examples so it learns a desired style, format, or behavior that prompting alone cannot reliably produce.

Adapting a model by training a small set of added or selected weights while freezing most parameters, cutting compute and storage costs compared with full fine-tuning.

The loss of a model's general capabilities when it is trained too narrowly or aggressively on a specialized dataset, showing up as degraded performance outside the training task.

Selecting, cleaning, deduplicating, and labeling examples for adaptation, since data quality and representativeness usually matter more than raw volume for downstream results.

A held-out, representative set of inputs with expected outputs used to measure model quality objectively and to compare prompts, models, or fine-tuned versions.

Scoring model outputs programmatically using metrics, exact-match checks, or a judge model, enabling repeatable regression testing at scale across many candidate versions.

Routing selected model outputs to people for judgment or approval, used to validate automated scores and to gate high-impact actions before they take effect.

Ongoing measurement of live model quality and safety against baselines, catching drift after data, prompts, or model versions change in production.

Gradual degradation of model usefulness as real-world inputs, user behavior, or source data shift away from what the model was trained or evaluated on.

Granting each component and user only the permissions needed for its specific job, limiting the blast radius if a credential or application is compromised.

An identity with attached permissions that workloads assume temporarily to call model and data services, avoiding long-lived access keys embedded in application code.

Storing API keys and credentials in a managed vault with rotation and audit rather than in source code, configuration files, or prompt text where they can leak.

An attack where untrusted text, such as a retrieved document or user message, contains instructions that hijack the model into ignoring its intended task or leaking data.

Unauthorized removal of sensitive information, for example by persuading a model to include confidential context in its response or by sending data to an unapproved endpoint.

Detecting and masking personal identifiers in prompts, logs, and retrieved content so sensitive data is not stored, transmitted, or echoed back by the model.

A configurable control that inspects inputs or outputs and blocks, filters, or rewrites content that violates safety, policy, or topical constraints before users see it.

Recording requests, prompts, retrieved sources, token counts, and latencies so teams can debug failures, audit behavior, and attribute cost to specific features or tenants.

Tagging and aggregating token usage and infrastructure spend by application, team, or customer so chargeback is accurate and expensive usage patterns become visible.

Pre-purchased model capacity that guarantees consistent performance under load, chosen over pay-per-token pricing when traffic is steady and latency predictability matters.

Reusing prior model responses or computed embeddings for repeated queries, cutting both latency and token spend when the same inputs recur frequently.

Documentation describing a model's intended uses, training approach, evaluation results, and known limitations, supporting informed selection and governance review.