AI Agent
An AI system that can plan steps, call tools, read or write data, and report results against a goal. Enterprise use needs permissions, logs, failure handling, and human review.
An AI system that can plan steps, call tools, read or write data, and report results against a goal. Enterprise use needs permissions, logs, failure handling, and human review.
The management framework for how AI may be used, which data it can access, who approves it, how risk is reviewed, and who is accountable for outcomes.
The ability of a team to understand what AI can and cannot do, how to use it safely, and where human judgment is still required.
A readiness review of workflows, data, systems, permissions, people, and governance before turning AI ideas into production projects.
The practice of identifying and controlling risks such as wrong answers, data leakage, bias, compliance gaps, model cost, and operational instability.
A content and technical approach that helps search engines and AI answer systems understand, summarize, and cite a website more accurately.
A defined set of steps for solving a problem. AI models usually combine algorithms, data, parameters, and infrastructure.
An application programming interface that lets systems exchange data or call functions. LLM integration often starts with APIs but still needs permissions, logs, and workflow design.
A theoretical AI that can generalize across many tasks at or beyond human level. Most business projects today still focus on narrower task-specific AI.
A mechanism used by Transformer models to evaluate relationships between parts of the input. It is one of the core ideas behind modern LLMs.
Answer Engine Optimization. The practice of making content easier for answer engines, AI summaries, and question-answer interfaces to understand and cite.
Often used to mean AI Engine Optimization or AI search optimization, but the market definition is less settled than AEO. It is best explained together with AEO and SEO.
A skew in data, models, or processes that leads to unfair or inaccurate output. Teams reduce it through data checks, human review, and monitoring.
A conversational interface that answers users. Modern chatbots can connect LLMs, knowledge bases, support systems, and business workflows.
The process of splitting documents into smaller searchable pieces for RAG. Chunks that are too small lose context; chunks that are too large reduce precision.
A model task that assigns data to categories, such as case type, support topic, document subject, or risk level.
AI that interprets images or video, used for inspection, recognition, monitoring, and medical image support.
The amount of text a model can read and process at once. A larger window does not guarantee better answers; data structure and prompts still matter.
The practice of managing data sources, quality, permissions, versions, retention, and ownership so AI can be trusted.
Adding correct labels or answers to data so it can be used to train, evaluate, or improve models.
The flow that extracts, cleans, transforms, stores, and delivers data for analysis or AI use.
Protection of personal and sensitive information from unauthorized use, especially when AI systems process business or customer data.
A collection of data used for training, testing, evaluation, or retrieval. Data quality strongly affects AI quality.
Machine learning with multi-layer neural networks that learn patterns from large data sets, often used for language, image, and speech models.
A generative model family often used for images, video, or audio, creating output by gradually removing noise.
A numeric representation of text, images, or data that enables semantic search, similarity matching, and RAG retrieval.
Evaluation workflows that test AI output for correctness, stability, safety, and cost. Enterprise projects should keep test cases and pass criteria.
AI techniques and interfaces that make outputs or decisions easier for people to understand, especially in audited or high-risk workflows.
Additional training on top of an existing model so it better fits a specific task, tone, classification, or format.
A model trained on broad data that can be adapted to many downstream tasks such as text, image, audio, code, or multimodal workflows.
AI that can create text, images, code, summaries, audio, or video. Business use should include data, permissions, and review controls.
A graphics processing unit designed for parallel computation and widely used for training and running large AI models.
Controls that limit AI inputs, outputs, tool calls, and data access to reduce unsafe content, overreach, and errors.
A confident but unsupported or incorrect AI output. RAG, source citation, and human review can reduce this risk.
A workflow design that keeps people involved at review, approval, correction, or escalation points.
The process in which a model receives input and produces output. Teams must manage response time, cost, stability, and data protection.
A managed collection of documents, FAQs, rules, cases, and internal knowledge. It can support RAG with source-grounded answers.
A language model trained at large scale to understand and generate natural language for summarization, Q&A, classification, writing, translation, and workflow assistance.
The time between a user request and the system response. AI features often need to balance model capability, speed, and cost.
A field where systems learn patterns from data instead of relying only on manually written rules.
Data that describes other data, such as source, date, author, category, permission, or status. It is important for search, RAG, and governance.
Machine learning operations. The discipline of deploying, monitoring, versioning, testing, and rolling back models so AI systems remain stable.
A drop in model performance when data, behavior, or the environment changes over time.
Tracking quality, errors, latency, cost, output behavior, and anomalies after a model is deployed.
A routing layer that chooses models based on task, cost, speed, quality, or risk needs.
AI that can process more than one data type, such as text, images, audio, video, and tables.
Natural language processing. The field that helps computers understand, analyze, and generate human language.
A layered computational structure that learns relationships in data and forms the core of many deep learning models.
Coordinating models, tools, data sources, and workflow steps in the right order, often used in AI agents and enterprise automation.
A learned internal value in a model. Parameter count can affect capability and cost, but it is not the only measure of quality.
Personal data that can identify a person, such as name, phone, ID number, medical record, or address. AI systems must protect it carefully.
The instruction and context given to a model. Good prompts define goals, data, output format, and constraints clearly.
The practice of designing prompts, examples, constraints, and output formats so models complete tasks more reliably.
An attack that tries to make a model ignore its original rules or reveal information through malicious input.
Retrieval-Augmented Generation. The system first retrieves relevant content from data sources, then asks the LLM to answer based on that content.
Ordering results by relevance, importance, or score. It is used in search, recommendation, review, and retrieval workflows.
A system that suggests products, articles, services, or next actions based on behavior, content, or similarity.
A learning approach where behavior improves through rewards and feedback, often seen in control, games, robotics, and alignment research.
AI practices that emphasize fairness, transparency, safety, privacy, explainability, and accountability.
Finding relevant content from a database, search engine, or knowledge base. It is the key step before generation in RAG.
Search based on meaning and context rather than exact keywords, often powered by embeddings and vector databases.
Detecting positive, negative, or neutral emotion in text, useful for support, reviews, social media, and surveys.
A smaller language model that costs less and can be faster, useful for edge devices or narrow tasks.
Turning spoken audio into text for meeting notes, customer service, subtitles, and voice-operated systems.
Data with fixed fields and formats, such as forms, database tables, JSON-LD, and report columns.
Training a model with examples that include correct answers, often used for classification, prediction, and recognition.
Data generated by a system for testing, training, or privacy protection. It still requires quality checks.
A setting that controls randomness in model output. Lower values are more stable; higher values are more creative but less consistent.
The basic unit of text processed by a model. Token count affects cost, speed, and context capacity.
Letting an LLM call APIs, databases, search, calculations, or enterprise system functions when needed.
Data used to build or adjust a model. Quality, permission, and representativeness affect performance and risk.
A model architecture based on attention mechanisms. It is the foundation of modern LLMs and many generative AI systems.
Data without fixed fields, such as PDFs, emails, images, audio, video, and free text.
Learning from unlabeled data to find patterns, clusters, or features.
A database that stores embeddings and supports similarity search, often used for RAG and semantic search.
A vision-language model that understands both images and text for image Q&A, document understanding, screen analysis, and multimodal workflows.
Automating repeated steps, notifications, reviews, synchronization, and reports. AI can become a judgment or summary point inside the workflow.
Completing a task from its description without task-specific examples or training data.