watsonx GenAI Engineer key terms

The managed IBM Cloud service that deploys, hosts, and monitors models and prompt templates. It provides inference endpoints, deployment spaces, and scoring APIs that watsonx.ai applications call at runtime.

A collaborative WML workspace that groups deployments, assets, and members with shared access. It isolates resources by project or team and controls who can deploy, score, or administer the assets inside it.

The organizational containers in watsonx.ai. Projects hold assets for collaboration and experimentation, while the resource hub catalog shares promoted assets such as models, prompts, and notebooks across the account.

The execution environment that runs notebooks, training jobs, and deployment workloads with selectable software specifications and hardware sizes. It supplies the compute behind both experimentation and production inference.

The platform component for building AI assistants and automating multi-step business workflows. It connects skills, tools, and language models so an agent can call systems and complete tasks on a user's behalf.

The platform component that tracks model and application risk throughout the lifecycle. It records factsheets, evaluations, and approvals, and produces audit evidence aligned to internal policy and external regulation.

A large pretrained model, typically a transformer trained on broad unlabeled data, that can be adapted to many downstream tasks through prompting or tuning rather than task-specific training from scratch.

The browsable inventory of foundation models available in watsonx.ai, with details such as provider, context window, supported languages, and task suitability. Candidates use it to select a model before prompting or tuning.

The maximum number of tokens a model can process in a single request, counting both the prompt and the generated output. Exceeding it forces truncation, chunk reduction, or a model with a larger limit.

The unit of text a model reads and generates, roughly a word fragment rather than a whole word. Token counts drive context limits, latency, and cost estimates for inference requests.

The process of splitting raw text into tokens using a model-specific vocabulary. Mismatched tokenizers between training and inference, or unusual characters, can change token counts and model behavior.

A reusable prompt with variables that are filled at runtime, such as instructions, retrieved context, and the user question. It standardizes inputs and supports versioning and evaluation.

The practice of designing instructions, examples, and output constraints so a foundation model produces accurate, consistent, and safe responses for a specific task without changing model weights.

A technique that supplies a small number of worked examples inside the prompt so the model infers the desired format and reasoning pattern. It improves consistency for classification and extraction tasks.

Asking a model to perform a task using only instructions and no worked examples. It is cheapest in tokens but less reliable for unusual formats or domain-specific labeling schemes.

A prompting pattern that asks the model to reason step by step before giving a final answer. It can improve multi-step accuracy but increases token use and may expose reasoning that should not be shown to end users.

Settings that control generation. Temperature and top-p adjust randomness and diversity, while max tokens caps output length. Lower temperature favors deterministic answers for extraction and classification tasks.

An attack in which untrusted text, such as a retrieved document or user input, contains instructions that override the developer's prompt. Defenses include input isolation, instruction hierarchy, and output validation.

Configurable controls that filter or transform model input and output, such as blocking harmful content, masking personal data, or enforcing topic limits. They reduce risk but can add latency and false positives.

A content-safety filter that scores text for hate, abuse, and profanity so requests or responses can be blocked or flagged. Thresholds are tuned to balance user safety against over-blocking legitimate content.

The identification of personal data such as names, addresses, or account numbers in text. In generative pipelines it is used to mask inputs before sending them to a model and to sanitize outputs.

An architecture that retrieves relevant passages from an external knowledge source and inserts them into the prompt so the model answers from current, private evidence instead of relying only on pretrained knowledge.

A dense numeric vector that represents the meaning of text, produced by an embedding model. Similar meanings map to nearby vectors, enabling semantic search and clustering.

A store optimized for indexing and similarity search over embeddings at scale. It returns the nearest passages for a query vector and often stores metadata used for filtering.

A similarity measure equal to the cosine of the angle between two vectors, ranging from -1 to 1. It ignores magnitude, which suits normalized text embeddings used in semantic retrieval.

Retrieval ranked by meaning similarity between query and document embeddings rather than exact keyword matches. It handles paraphrases well but can miss precise identifiers that lexical search catches.

A retrieval strategy that combines dense semantic similarity with sparse keyword scoring such as BM25, then fuses the ranked results. It improves recall for both conceptual questions and exact terms.

An index-based method that finds close vectors quickly by trading a small amount of recall for large speed gains. HNSW is a common graph-based ANN algorithm used in vector stores.

Splitting source documents into passages small enough to embed and fit in the prompt. Chunk size and overlap strongly affect retrieval precision and the coherence of the retrieved context.

Restricting vector search to records matching attributes such as date, department, or access level. It improves relevance and enforces permissions so users retrieve only documents they may see.

A second retrieval stage that scores candidate passages with a more precise cross-encoder or model-based judge and reorders them before prompting. It raises answer quality at extra latency and cost.

Further training of a pretrained model on task-specific labeled examples so its weights adapt to a domain, style, or output format. It changes the model itself rather than only its prompt.

Supervised fine-tuning on prompt-response pairs that teach a model to follow instructions and desired formats. It is often a precursor to preference optimization for helpfulness and safety.

A parameter-efficient tuning method that freezes base weights and trains small low-rank matrices, greatly reducing memory and cost. Adapters can be swapped without duplicating the full model.

A post-training approach that trains a reward model from human preference comparisons and optimizes the language model against it. It shapes tone and helpfulness but requires careful data and evaluation.

Curating, cleaning, deduplicating, and formatting examples for tuning, including train-validation splits and prompt-response structure. Poor data quality, leakage, or imbalance undermines tuning gains regardless of hyperparameters.

A training setting chosen before training, such as learning rate, batch size, or number of epochs. These control convergence and overfitting and are tuned on validation data rather than learned from it.

Systematic measurement of model quality using task metrics, held-out data, and human review. In generative settings it covers accuracy, groundedness, safety, and latency rather than a single score.

The degree to which a generated answer is supported by the provided source context rather than invented. It is a core RAG quality metric and a frequent target of automated evaluation.

A confident model output that is factually wrong or unsupported by evidence. Mitigations include retrieval grounding, citation requirements, lower temperature, and validation of structured outputs.

A structured record in watsonx.governance describing a model's purpose, training data, evaluations, and limitations. It supports review, transparency, and audit reporting across the model lifecycle.

Documentation summarizing a model's intended use, training approach, performance, and known limitations. It informs selection and risk assessment and feeds governance records.

The discipline of identifying, measuring, mitigating, and monitoring risks from models, including bias, drift, and misuse. It assigns ownership, sets controls, and requires periodic review.

A governed sequence of reviews and sign-offs that a model or use case must pass before production. It records approvers, evidence, and conditions, and can block deployment until requirements are met.

The documented path of data from source through transformation to model input. It supports reproducibility, audit, and impact analysis when upstream data changes or is disputed.

Replacing sensitive values with realistic substitutes or redaction so development and testing can use data without exposing personal information. It reduces privacy risk in generative pipelines.

An orchestration pattern where a language model plans steps, calls tools, observes results, and iterates toward a goal. It requires guardrails, retries, and human checkpoints for high-impact actions.

A model capability to emit structured requests for external functions, such as database lookups or ticketing actions, whose results are fed back into the conversation. It connects language reasoning to real systems.

An open-source framework for composing prompts, models, retrievers, and tools into chains and agents. It is commonly used with watsonx.ai models to build RAG and agent applications.

A deployed HTTP interface that returns model predictions or generations for live requests. It is configured with a serving environment, scaling policy, and authentication, and is monitored for latency and errors.

Degradation in model performance as real-world data diverges from training or evaluation conditions. Monitoring detects it through input distributions, quality metrics, and feedback so models can be retrained or retired.

A stored, versioned artifact in watsonx.ai representing a trained model or prompt configuration that can be evaluated, promoted, and deployed. It carries metadata used by governance and deployment.