Glossary
The shared language for modern AI builders
Browse clear definitions of the concepts, protocols, tools, standards, and terminology shaping agentic systems, developer tooling, evaluation, and the broader AI ecosystem as it stands today.
No terms match your search.
A
- A/B DeploymentA deployment strategy that routes production traffic between two application versions to compare performance, reliability, or user outcomes before full rollout.
- A/B TestingAn experimentation method that compares two variants under similar conditions to determine which performs better against predefined metrics.
- Abductive ReasoningARA reasoning process that infers the most likely explanation for observed facts or evidence, forming plausible hypotheses from incomplete information.
- AccelerateAn open-source Hugging Face library that simplifies distributed training and inference across CPUs, GPUs, TPUs, and multiple devices with minimal code changes.
- AcceleratorA specialized hardware device designed to speed up compute-intensive workloads such as AI training and inference by efficiently executing parallel operations.
- Acceptance TestingA testing process that verifies whether an AI system satisfies specified requirements and is ready for deployment or release.
- Access TokenA credential issued after authentication that authorizes a client to access protected resources or APIs for a limited period.
- AccountabilityThe principle that individuals and organizations are responsible for the decisions, actions, and outcomes of AI systems throughout their lifecycle.
- AccuracyA metric that measures the proportion of correct predictions or outputs produced by a model or AI system out of all evaluated cases.
- Action ExecutionThe process of carrying out a selected action or tool invocation as part of an agent's plan while tracking its outcome and execution state.
- Action SelectionThe process of choosing the most appropriate action from available options based on goals, policies, context, constraints, or expected outcomes.
- Active LearningA machine learning approach in which a model selectively requests labels for the most informative data samples to improve performance with minimal annotation effort.
- Adaptation PolicyA set of rules or strategies that governs how an AI agent adjusts its behavior, plans, or parameters in response to changing conditions or feedback.
- AdapterA software component or pattern that enables incompatible systems, interfaces, or protocols to work together by translating requests, responses, or data formats.
- Adaptive AutonomyThe ability of an AI system to dynamically adjust its level of independence based on context, objectives, risks, or human input.
- Adversarial TestingA testing method that deliberately uses challenging, deceptive, or malicious inputs to evaluate an AI system's robustness, reliability, and security.
- Agent CatalogA structured collection of registered AI agents and their metadata, enabling discovery, selection, and management based on capabilities and attributes.
- Agent CollaborationThe coordinated interaction of multiple AI agents that communicate, share responsibilities, and work together to achieve a common objective.
- Agent CommunicationThe exchange of messages, data, or requests between AI agents to coordinate actions, share information, and accomplish collaborative tasks.
- Agent ContractA formal agreement that defines the capabilities, responsibilities, communication rules, and expectations governing interactions between AI agents.
- Agent ConversationA structured exchange of messages between AI agents that enables coordination, information sharing, negotiation, or collaborative task execution.
- Agent CoordinationThe process of organizing and synchronizing multiple AI agents to achieve shared goals through communication, task allocation, and cooperative decision-making.
- Agent DirectoryA searchable directory that lists available AI agents along with their capabilities, metadata, and endpoints to support discovery and selection.
- Agent GuardrailsPolicies, constraints, and runtime controls that keep an AI agent operating within defined safety, security, and behavioral boundaries.
- Agent MetadataStructured information that describes an AI agent's identity, capabilities, endpoints, configuration, and other attributes to support discovery and interoperability.
- Agent NegotiationA process in which AI agents communicate to reach agreements on tasks, resources, responsibilities, or decisions based on shared goals and constraints.
- Agent RegistryA centralized repository that stores AI agent metadata and capabilities, enabling agents and applications to discover, identify, and interact with available agents.
- Agent RoutingThe process of directing requests, messages, or tasks to the most appropriate AI agent based on capabilities, availability, policies, or context.
- Agent SessionA stateful interaction context that maintains conversation history, shared context, and execution state between communicating AI agents.
- Agent TestingThe process of evaluating an AI agent's behavior, decision-making, tool use, and task execution to verify correctness, reliability, and robustness.
- Agent WorkflowA structured sequence of tasks, decisions, and tool interactions that an AI agent follows to achieve a specific goal or complete a process.
- Agent-to-AgentA2AA communication paradigm that enables AI agents to exchange information, coordinate actions, negotiate tasks, and collaborate using standardized protocols.
- AgentOpsAn observability platform for monitoring, tracing, debugging, and evaluating AI agents and agentic workflows throughout development and production.
- AgnoAn open-source framework for building AI agents and multi-agent applications with support for memory, tool integration, knowledge retrieval, and workflow orchestration.
- AI AgentA software system that perceives its environment, reasons about goals, and autonomously performs actions using AI models, tools, memory, and workflows.
- AI Coding AssistantAn AI-powered development tool that helps write, explain, review, refactor, and debug code by understanding natural language and programming context.
- AI GovernanceThe framework of policies, processes, and oversight that ensures AI systems are developed, deployed, and operated responsibly, safely, and in compliance with regulations.
- AI IDEAn integrated development environment enhanced with AI capabilities that assists developers with code generation, debugging, navigation, refactoring, and project understanding.
- AI ObservabilityThe practice of monitoring, tracing, and analyzing AI systems to understand their behavior, performance, reliability, and operational health in production.
- AI Risk ManagementThe process of identifying, assessing, mitigating, and monitoring risks associated with the development, deployment, and operation of AI systems.
- AlertingThe process of automatically notifying users or systems when predefined conditions, thresholds, or anomalies indicate potential issues requiring attention.
- AllowlistA predefined list of approved inputs, entities, actions, or resources that an AI system explicitly permits while rejecting all others.
- AMQPAMQPAn open standard messaging protocol that enables reliable, interoperable communication between distributed applications through message brokers and queues.
- Anomaly DetectionThe process of identifying unusual patterns, behaviors, or events that deviate from expected system behavior and may indicate failures, risks, or performance issues.
- AnthropicAn AI research and technology company that develops Claude language models, APIs, and tools for building safe and reliable generative AI applications.
- Anthropic SDKThe official software development kit from Anthropic for building applications that interact with Claude models through the Anthropic API.
- Anti-Corruption LayerACLAn architectural pattern that isolates a system from incompatible external models by translating data, behaviors, and interfaces between them.
- Apache AvroAn open-source data serialization framework that uses compact binary encoding and schemas for efficient, interoperable data exchange between systems.
- Apache ParquetAn open-source columnar data storage format optimized for efficient compression, analytics, and interoperability across big data and AI processing systems.
- API ClientA software library or component that simplifies communication with an API by handling requests, authentication, responses, and error handling.
- API ExecutionThe process of invoking an external API to perform an action, retrieve data, or interact with a service as part of an AI agent's workflow.
- API GatewayA service that acts as a single entry point for APIs, routing requests while handling authentication, rate limiting, monitoring, and other cross-cutting concerns.
- API KeyA unique secret credential used to authenticate an application or client when accessing an API, typically identifying the caller and authorizing requests.
- API TestingThe process of verifying that an API functions correctly, reliably, securely, and as expected by validating its requests, responses, behavior, and error handling.
- API WrapperA software layer that simplifies interactions with an API by abstracting request handling, authentication, data transformation, and error management.
- Application MonitoringThe practice of collecting and analyzing telemetry from applications to track performance, availability, errors, and overall operational health.
- Approval GateA designated checkpoint where human approval is required before an AI system can proceed with an action, workflow, or decision.
- ARCARCA benchmark that evaluates an AI model's ability to solve grade-school science questions requiring reasoning, knowledge, and commonsense understanding.
- Arena-HardA benchmark that evaluates advanced language models using difficult, real-world prompts and pairwise comparisons to measure instruction-following and overall response quality.
- Arize PhoenixAn open-source AI observability platform for tracing, evaluating, and debugging LLM applications, retrieval pipelines, and agentic workflows.
- Assistant AgentAn AI agent designed to assist users by understanding requests, providing information, and performing tasks through conversation, tools, and automation.
- Assistant PromptA message generated by an AI assistant during a conversation, representing the model's response and serving as context for subsequent interactions.
- Assistants APIAn API that enables developers to build AI assistants with persistent conversations, tool use, file handling, and multi-step task execution.
- AsyncAPIAn open specification for designing, documenting, and managing event-driven and asynchronous APIs across messaging systems and streaming platforms.
- Atomic AgentsAn open-source framework for building modular AI applications using reusable, composable agents with strongly typed inputs, outputs, and workflows.
- Attribute-Based Access ControlABACAn access control model that grants or denies access based on attributes of users, resources, actions, and environmental conditions.
- Audit TrailA chronological record of actions, events, and system changes that enables traceability, accountability, compliance, and forensic analysis.
- AuthenticationThe process of verifying the identity of a user, application, or system before granting access to protected resources or services.
- AuthorizationThe process of determining what actions or resources an authenticated user, application, or system is permitted to access.
- AutoGenAn open-source framework from Microsoft for building AI agents and multi-agent systems that collaborate through conversations, tools, and automated workflows.
- Automated EvaluationAn evaluation method that uses software, benchmarks, metrics, or models to assess the quality, correctness, or performance of AI systems without manual review.
- Autonomous AgentAn AI agent that independently perceives, reasons, plans, and executes actions to achieve goals with minimal or no human intervention.
- Autonomous ExecutionThe ability of an AI system to execute tasks and decisions independently while operating within defined goals, constraints, and safety policies.
- Autonomy LevelThe degree of independence an AI system has in making decisions and performing actions without requiring human intervention or approval.
- AutoscalerA system that automatically adjusts computing resources based on workload demand to maintain performance, availability, and efficient resource utilization.
- AutoscalingThe automatic adjustment of computing resources in response to changing workloads to optimize performance, availability, and cost efficiency.
- Azure AI SearchA cloud-based search service from Microsoft that provides full-text, vector, and hybrid search capabilities for AI applications and retrieval-augmented generation workflows.
- Azure OpenAI SDKA software development kit for integrating Azure OpenAI services into applications using Azure authentication, deployment, and management features.
B
- Bare Metal ServerA physical server dedicated to a single tenant, providing direct access to hardware resources without a virtualization layer.
- Batch APIAn API that processes multiple requests asynchronously in a single job, improving throughput and reducing costs for large-scale workloads.
- Batch InferenceThe process of running model inference on multiple inputs together to improve throughput, resource utilization, and overall serving efficiency.
- Bayesian OptimizationAn optimization method that uses probabilistic models to efficiently identify promising parameter configurations with minimal evaluation runs.
- Bearer TokenA security token that grants access to protected resources solely by possession, without requiring additional proof of identity.
- Benchmark RunA single execution of a benchmark used to measure and record a model or system's performance under defined conditions.
- BERTScoreBERTScoreA text generation metric that measures semantic similarity using contextual embeddings from BERT instead of exact token overlap.
- Bias TestingThe process of evaluating an AI system for unfair, systematic, or discriminatory behavior across different users, groups, or scenarios.
- Binary ProtocolA communication protocol that encodes data in a compact binary format to improve transmission efficiency and processing performance.
- Blackboard ArchitectureAn AI architecture where multiple specialized agents collaborate by reading from and writing to a shared knowledge space to solve complex problems.
- BLEUBLEUA metric that evaluates generated text by measuring n-gram overlap between a candidate output and one or more reference texts.
- Blind EvaluationAn evaluation method where evaluators assess outputs without knowing which model, system, or approach produced them to reduce bias.
- Block StorageA storage system that organizes data into fixed-size blocks, providing low-latency, high-performance storage for operating systems and applications.
- Blue-Green DeploymentA deployment strategy that uses two identical production environments, allowing traffic to switch from the current version to a new one with minimal downtime.
- BoilerplateA reusable starter codebase or project template that provides a predefined structure and common functionality for faster development.
- Bounded AutonomyAn approach to autonomy where an AI system operates independently only within predefined goals, constraints, permissions, or safety limits.
- BraintrustAn AI evaluation and observability platform for testing, tracing, benchmarking, and improving the quality of LLM applications.
- BridgeAn interoperability pattern that connects incompatible systems by translating or forwarding communication between their interfaces or protocols.
- Broadcast MessagingA communication pattern where a message is sent simultaneously to multiple agents or recipients without targeting a specific destination.
- Broker PatternAn integration pattern where a broker mediates communication between distributed systems, routing messages and decoupling senders from receivers.
- BSONBSONA binary serialization format for JSON-like documents that supports efficient storage, transmission, and additional data types.
- Build ToolA software tool that automates compiling, packaging, testing, and other build tasks required to produce deployable applications.
C
- CacheA storage layer that temporarily keeps frequently accessed data in fast memory to reduce latency and improve application performance.
- Camel AICAMELAn open-source framework for building, simulating, and orchestrating collaborative AI agents that communicate and solve tasks together.
- Canary DeploymentA deployment strategy that gradually releases a new application version to a small subset of users before expanding it to everyone.
- Canary TestingA testing approach that exposes changes to a small subset of users or traffic to validate behavior before a full rollout.
- Canonical Data ModelCDMA standardized data model that enables different systems to exchange information through a common, shared representation.
- Capability AdvertisementThe process of publishing an agent's available capabilities, interfaces, or services so they can be discovered and invoked by other systems.
- Capability BindingThe process of associating a declared capability with a specific tool, service, or implementation so it can be invoked by compatible systems.
- Capability DescriptorA structured description of an agent's capabilities, interfaces, requirements, and metadata used for discovery and interoperability.
- Capability DiscoveryThe process of identifying available agent or service capabilities so they can be selected and invoked by compatible systems.
- Capability ExchangeThe process by which agents communicate and share their supported capabilities to discover compatible collaborators and coordinate tasks.
- Capability RegistryA registry that stores and organizes capability metadata, enabling agents and services to discover available functions and interfaces.
- Causal ReasoningA reasoning process that identifies cause-and-effect relationships between events or variables to support prediction, explanation, and decision-making.
- Chain of ThoughtCoTA reasoning technique in which an AI model generates intermediate reasoning steps before producing a final answer or decision.
- Chat Completions APIAn API for generating conversational AI responses by sending a sequence of chat messages to a language model.
- Chatbot ArenaA crowdsourced benchmark where users compare responses from multiple AI models through blind pairwise voting to rank model performance.
- CheckpointingThe practice of saving an execution state at defined points so a task or workflow can resume from the last checkpoint after interruption or failure.
- ChromaAn open-source vector database designed for storing, indexing, and retrieving embeddings to power semantic search and retrieval-augmented applications.
- Chunk OverlapThe practice of sharing content between adjacent document chunks to preserve context and improve retrieval accuracy across chunk boundaries.
- Circuit BreakerA resilience pattern that temporarily stops requests to a failing service to prevent cascading failures and allow recovery.
- Citation RetrievalThe process of retrieving source documents or passages that support generated responses with verifiable references and evidence.
- Citation VerificationThe process of validating that cited sources are accurate, relevant, and correctly support the claims made in an AI-generated response.
- CLICLIA command-line interface that enables users to interact with software by entering text commands instead of using a graphical interface.
- CloudEventsAn open specification for describing event data in a consistent format, enabling interoperable event-driven systems across platforms and services.
- ClusterA group of interconnected computers that work together as a single system to provide scalable computing, storage, or application services.
- Code GenerationThe automated creation of source code from natural language, specifications, prompts, or structured inputs using AI or software tools.
- Code GeneratorA tool that automatically produces source code, project files, or boilerplate from templates, specifications, or structured inputs.
- Code ReviewThe practice of examining source code to identify defects, improve quality, ensure maintainability, and verify compliance with coding standards.
- Coding AgentAn AI agent specialized in writing, debugging, testing, refactoring, and maintaining software by autonomously performing coding tasks.
- CohereAn AI company that develops large language models, embedding models, and enterprise AI platforms for generative and retrieval applications.
- Cohere SDKA software development kit for integrating Cohere's language, embedding, reranking, and AI generation models into applications.
- Collaboration StrategyA structured approach that defines how multiple agents cooperate, share responsibilities, communicate, and coordinate to achieve shared objectives.
- Comparative EvaluationAn evaluation method that compares the performance of two or more models, systems, or approaches using the same tasks and criteria.
- Compatibility LayerAn abstraction layer that enables incompatible systems, APIs, or software components to work together without requiring direct modifications.
- CompletenessA metric that measures how fully an AI response covers the required information, tasks, or expected outputs for a given request.
- ComplianceThe practice of ensuring AI systems, processes, and data handling conform to applicable laws, regulations, standards, and organizational policies.
- Compute InstanceA virtual or physical compute resource that provides CPU, memory, storage, and networking for running applications and workloads.
- Concept DriftA change in the relationship between inputs and expected outputs over time, causing a model's predictions to become less accurate.
- Conditional WorkflowA workflow that dynamically selects different execution paths based on conditions, rules, events, or intermediate results.
- Confidence IntervalCIA statistical range that estimates where a true value is likely to lie with a specified level of confidence based on sampled data.
- Confidence ScoringThe process of estimating how certain an AI system is about a prediction, decision, or generated response using a confidence value.
- Configuration FileA file that stores settings, parameters, and options used to configure the behavior of an application, tool, or system.
- ConnectorA software component that enables communication and data exchange between different systems, services, applications, or external resources.
- ConsensusA coordination process in which multiple agents reach a shared decision or agreement before taking actions or progressing a workflow.
- Consent ManagementThe process of obtaining, recording, updating, and enforcing user consent for collecting, processing, and sharing personal data.
- ConsistencyA metric that measures how reliably an AI system produces stable, coherent, and similar outputs for equivalent inputs or repeated evaluations.
- Constraint SatisfactionThe process of selecting actions or solutions that satisfy a defined set of rules, limitations, requirements, or operational constraints.
- Container RegistryA repository for storing, managing, versioning, and distributing container images used in application deployment workflows.
- ContainerizationThe practice of packaging an application and its dependencies into portable containers that run consistently across different environments.
- Content Delivery NetworkCDNA geographically distributed network of servers that caches and delivers content with low latency and high availability to users.
- Content EncodingA mechanism that specifies how message content is compressed or transformed for transmission and how it should be decoded by the recipient.
- Content FilteringThe process of detecting, blocking, or modifying content that violates safety policies, compliance requirements, or application rules.
- Content NegotiationA protocol mechanism that enables clients and servers to agree on the most appropriate representation, format, or encoding for exchanged content.
- Context InjectionThe process of supplying relevant context to a model or agent at runtime to improve task execution and response quality.
- Context OptimizationThe process of improving the selection, organization, and delivery of context to maximize model accuracy, relevance, and efficiency.
- Context ProviderA component that supplies relevant context, data, or resources to models and agents during execution to support accurate task completion.
- Context ResolutionThe process of identifying, retrieving, and assembling the most relevant context needed by a model or agent for a specific request.
- Context SerializationThe process of converting context into a structured format for storage, transmission, or exchange between models, agents, or systems.
- Context ServerA server that provides, manages, and delivers contextual data, resources, or tools to models and agents during runtime.
- Context StoreA repository that stores contextual information so it can be retrieved, updated, and shared across model or agent interactions.
- Context SynchronizationThe process of keeping contextual information consistent and up to date across models, agents, sessions, or distributed systems.
- Context TransferThe process of passing contextual information between models, agents, sessions, or systems to preserve continuity and support coordinated execution.
- Context WindowThe maximum amount of information a model can process and retain within a single interaction or inference.
- Contextual PromptingA prompting technique that incorporates relevant context into a prompt to improve the accuracy, relevance, and coherence of model responses.
- Continual LearningA learning approach in which an AI system continuously acquires new knowledge and skills over time while retaining previously learned capabilities.
- Continuous DeliveryCDA software delivery practice that automates building, testing, and preparing applications for reliable release to production on demand.
- Continuous EvaluationAn evaluation approach that continuously measures AI system performance throughout development and production to detect regressions and ensure quality.
- Continuous ImprovementAn ongoing process of using evaluation results, feedback, and operational insights to iteratively improve AI system performance and quality.
- Continuous IntegrationCIA development practice that frequently merges code changes into a shared repository where automated builds and tests verify software quality.
- Control PlaneThe component of a distributed system that manages configuration, coordination, routing policies, and operational control of services.
- Conversation ThreadA structured sequence of messages exchanged between agents that preserves conversational context and supports coordinated interactions over time.
- Coordination ProtocolA defined set of rules and message flows that enables multiple agents to coordinate actions, exchange information, and achieve shared goals.
- CorrectnessA metric that measures whether an AI system's output is factually accurate, logically valid, and satisfies the intended task or expected result.
- Cost MonitoringThe continuous tracking and analysis of infrastructure, API, and model usage costs to optimize spending and detect unexpected expenses.
- Cost OptimizationThe process of reducing infrastructure, model, and operational costs while maintaining or improving application performance, reliability, and quality.
- Cost per RequestA metric that measures the average monetary cost incurred to process a single API call, inference, or user request.
- CPUCPUThe central processing unit that executes program instructions and performs the general-purpose computations required by software applications.
- Credential StoreA secure storage system for managing credentials such as API keys, passwords, tokens, and certificates used by applications and services.
- Credential VaultA secure system for storing, managing, and controlling access to sensitive credentials such as passwords, API keys, certificates, and tokens.
- CrewAIAn open-source framework for building, orchestrating, and managing collaborative AI agents that work together to complete complex tasks.
- Cross EncoderA neural reranking model that jointly encodes a query and a candidate document to produce a highly accurate relevance score.
- CSVCSVA plain-text file format that stores tabular data as rows of values separated by commas or other delimiters.
- CUDACUDAA parallel computing platform and programming model developed by NVIDIA for accelerating applications using GPUs.
- Customer Support AgentAn AI agent specialized in assisting customers by answering questions, resolving issues, providing guidance, and automating support interactions.
D
- DashboardA visual interface that displays real-time and historical metrics, logs, and operational insights to monitor the health and performance of AI systems.
- Data AgentAn AI agent specialized in collecting, processing, analyzing, and managing data to generate insights, answer queries, and support decision-making.
- Data ContractA formal specification that defines the structure, format, semantics, and validation rules for data exchanged between systems or services.
- Data CurationThe process of collecting, organizing, cleaning, validating, and maintaining datasets to improve the quality of AI training, evaluation, and feedback pipelines.
- Data DeserializationThe process of converting serialized data from a structured format into in-memory objects or data structures that an application can process.
- Data DriftA change in the statistical distribution or characteristics of input data over time that can reduce the performance of an AI model.
- Data GovernanceThe framework of policies, processes, roles, and controls that ensures data is managed securely, consistently, and responsibly throughout its lifecycle.
- Data PlaneThe component of a distributed system that handles the execution, routing, and processing of application traffic and data according to policies defined by the control plane.
- Data ResidencyThe requirement that data be stored, processed, and managed within specific geographic locations or legal jurisdictions to satisfy regulatory, contractual, or organizational obligations.
- Data SchemaA structured specification that defines the format, fields, data types, relationships, and validation rules for exchanged data.
- Data SerializationThe process of converting data structures or objects into a structured format for storage, transmission, or exchange between systems.
- Decision AutonomyThe degree to which an AI agent or system can make decisions independently without requiring human approval or intervention.
- Decision PolicyA strategy or set of rules that determines how an AI agent selects actions or makes decisions based on goals, inputs, constraints, and current state.
- Deductive ReasoningDRA reasoning process in which an AI system derives logically necessary conclusions from general rules, facts, or premises.
- Delegated AutonomyA level of autonomy in which an AI agent is authorized by a human or another system to make decisions and execute tasks within predefined goals, permissions, and constraints.
- DeliberationA reasoning process in which an AI system evaluates multiple options, consequences, and trade-offs before selecting a decision or course of action.
- Dense RetrievalA retrieval technique that uses dense vector embeddings to identify semantically relevant documents based on similarity in an embedding space.
- DenylistA list of prohibited inputs, outputs, entities, patterns, or actions that an AI system blocks or rejects to enforce safety, security, or policy requirements.
- Dependency LockfileA file that records the exact versions and resolved dependencies of a project to ensure reproducible and consistent builds across environments.
- Dependency ManagementThe process of declaring, installing, updating, and maintaining the external libraries and packages required by a software project.
- Deployment ManifestA declarative configuration file that specifies how an application or service should be deployed, configured, and managed within a target environment.
- Deployment PipelineAn automated workflow that builds, tests, validates, and deploys applications or AI systems across development, staging, and production environments.
- DifyAn open-source platform for developing, deploying, and managing AI applications with visual workflow orchestration, prompt management, knowledge retrieval, and support for multiple large language models.
- Direct Preference OptimizationDPOA preference optimization technique that directly trains a language model to prefer chosen responses over rejected ones without requiring an explicit reward model.
- Discovery ProtocolA protocol that defines how agents or services advertise their capabilities, discover available resources, and retrieve metadata needed for interaction.
- Distributed ComputingA computing approach in which processing tasks are distributed across multiple interconnected computers that work together as a single system.
- Distributed InferenceAn inference approach that distributes model execution across multiple machines or accelerators to improve throughput, reduce latency, or support models that exceed the resources of a single device.
- Distributed TracingAn observability technique that tracks requests as they flow across multiple distributed services to measure latency, diagnose failures, and analyze system behavior.
- Distributed TrainingA training approach that distributes model training across multiple machines or accelerators to reduce training time and enable larger models or datasets.
- DockerAn open-source containerization platform for building, packaging, distributing, and running applications as isolated containers across different environments.
- Docker ComposeA tool and specification for defining, configuring, and running multi-container applications using a declarative configuration file.
- Docker ImageA portable, immutable package containing an application, its runtime, libraries, dependencies, and configuration needed to run consistently across environments.
- Document ChunkingThe process of splitting documents into smaller, semantically meaningful segments to improve indexing, retrieval, and context generation for AI systems.
- Document LoaderA component that ingests documents from various sources and formats, extracts their content, and prepares them for indexing and retrieval pipelines.
- DocVQADocVQAA benchmark for evaluating an AI model's ability to answer questions by understanding and extracting information from document images.
- Domain AgentAn AI agent specialized in a specific domain of knowledge or expertise, enabling it to perform domain-specific tasks with greater accuracy and contextual understanding.
- Drift DetectionThe process of identifying significant changes in data distributions, model behavior, or prediction patterns that may indicate degraded AI system performance.
- DSPyDSPyAn open-source framework for programming language model applications using declarative modules and automated prompt and parameter optimization.
- Dynamic PlanningA planning approach in which an AI system continuously updates or revises its plans in response to changing goals, environmental conditions, or new information.
E
- Edge DeploymentA deployment approach in which AI models or applications run on devices or servers located close to where data is generated or consumed to reduce latency and bandwidth usage.
- Elasticsearch Vector SearchA vector search capability in Elasticsearch that enables semantic search by indexing and querying vector embeddings using approximate nearest neighbor (ANN) algorithms.
- Embedding GenerationThe process of converting text, images, or other data into dense numerical vector representations that capture semantic meaning for retrieval, similarity search, and downstream AI tasks.
- Embedding ModelA machine learning model that converts text, images, audio, or other data into dense numerical vector representations that capture semantic meaning for retrieval, similarity search, and related AI tasks.
- Embeddings APIAn API that generates vector embeddings from text, images, or other supported inputs for use in semantic search, retrieval, clustering, recommendation, and related AI applications.
- End-to-End TestingE2EA testing approach that validates an entire application or AI workflow by exercising complete user scenarios across all integrated components and external dependencies.
- Environment VariablesNamed configuration values provided by the operating system or runtime environment that allow applications to access settings, credentials, and other configuration without hardcoding them into source code.
- Episodic MemoryEMA type of memory that stores information about specific events, experiences, and interactions, enabling an AI system to recall past episodes and use them to inform future reasoning and decision-making.
- Error AnnotationThe process of labeling, categorizing, and documenting errors in AI system outputs to support evaluation, debugging, model improvement, and feedback-driven training.
- Error RecoveryThe process by which an AI system detects execution failures, handles exceptions, and restores normal operation through retries, fallbacks, alternative actions, or corrective procedures.
- Error TrackingThe process of capturing, recording, aggregating, and monitoring errors and exceptions generated by an AI system to support debugging, reliability, and operational visibility.
- EscalationThe process of transferring a task, decision, or interaction from an AI system to a human or another specialized agent when predefined conditions, limitations, or risks are encountered.
- Escalation PolicyA set of predefined rules and procedures that determine when and how an AI system should transfer a task, decision, or incident to a human or higher-authority system.
- Evaluation AgentAn AI agent specialized in assessing the quality, accuracy, safety, or effectiveness of AI systems, workflows, or outputs using predefined metrics, evaluation criteria, or benchmarks.
- Evaluation DatasetA curated collection of test examples, inputs, and expected outcomes used to measure the quality, accuracy, safety, and reliability of AI models and systems.
- Event BrokerA messaging component that receives, routes, and distributes events between producers and consumers, enabling asynchronous communication in distributed systems.
- Event BusA communication infrastructure that transports and distributes events between producers and consumers, enabling loosely coupled, event-driven interactions across distributed systems.
- Event IntegrationThe integration of AI systems with event sources and event consumers, enabling applications to publish, receive, and respond to events across external services and internal workflows.
- Event-Driven ArchitectureEDAA software architecture pattern in which components communicate by producing, consuming, and reacting to events asynchronously, enabling loosely coupled, scalable, and resilient systems.
- Event-Driven WorkflowA workflow in which tasks and actions are initiated in response to events, allowing processes to execute asynchronously as new events occur.
- Exact MatchEMAn evaluation metric that measures whether a model's prediction exactly matches the expected reference answer, without allowing partial credit.
- Execution AgentAn AI agent specialized in executing tasks, invoking tools, carrying out plans, and interacting with external systems to achieve specified goals.
- Execution TraceA chronological record of the steps, decisions, tool invocations, inputs, outputs, and state transitions that occur during the execution of an AI system, agent, or workflow.
- Experience ReplayA learning technique in which an AI system stores past experiences and reuses them during training to improve learning efficiency, stability, and sample utilization.
- Experiment TrackingThe process of recording and managing information about AI experiments, including datasets, models, hyperparameters, code versions, metrics, and outcomes to enable reproducibility and comparison.
- Experimental ControlThe practice of keeping variables, conditions, and procedures consistent across experiments so that observed differences in outcomes can be attributed to the factors being tested.
- Expert ReviewAn evaluation method in which subject matter experts assess the quality, accuracy, safety, or effectiveness of an AI system, model, or output using their domain knowledge and established criteria.
- ExplainabilityXAIThe ability of an AI system to provide understandable reasons or evidence for how it produced a prediction, recommendation, or decision, enabling humans to interpret and trust its behavior.
- Explicit FeedbackFeedback that is intentionally provided by users or human evaluators through ratings, preferences, corrections, or written comments to assess or improve an AI system's outputs or behavior.
- ExtensionA software component that adds new features or capabilities to an existing application, platform, or development environment without modifying its core functionality.
- Extension PointA predefined interface or integration hook that allows additional functionality, custom logic, or third-party components to be added to a system without modifying its core implementation.
- External ActionsOperations performed by an AI system that interact with external tools, services, APIs, or physical systems to retrieve information, modify state, or execute tasks beyond the model's internal reasoning.
F
- F1 ScoreF1An evaluation metric that measures the balance between precision and recall by calculating their harmonic mean, providing a single score for classification and information retrieval performance.
- FacadeAn interoperability pattern that provides a simplified, unified interface to a more complex system, service, or collection of components, reducing integration complexity and hiding underlying implementation details.
- Fact CheckingThe process of verifying the factual accuracy of AI-generated content by comparing claims against trusted sources, evidence, or authoritative knowledge before or after a response is produced.
- Failure AnalysisThe process of investigating errors, failures, or unexpected behavior in an AI system to identify root causes, assess their impact, and implement corrective or preventive actions.
- FAISSFAISSAn open-source library developed by Meta for efficient similarity search and clustering of dense vector embeddings, commonly used to build high-performance semantic search and retrieval systems.
- FaithfulnessAn evaluation metric that measures whether an AI system's output is fully supported by the provided input, retrieved context, or source material without introducing unsupported or fabricated information.
- Feature FlagA configuration mechanism that enables or disables specific application features at runtime without requiring code changes or redeployment, allowing controlled rollouts, testing, and experimentation.
- FederationAn interoperability pattern in which multiple independent systems, organizations, or services collaborate under shared protocols and standards while retaining their own governance, control, and autonomy.
- Feedback AggregationThe process of collecting, combining, and organizing feedback from multiple users, evaluators, or automated systems to identify trends, measure performance, and guide AI system improvements.
- Feedback CollectionThe process of gathering feedback from users, human evaluators, or automated systems to assess AI system performance and provide data for evaluation, improvement, and model refinement.
- Feedback PipelineA structured workflow for collecting, processing, analyzing, and incorporating feedback into the evaluation, improvement, and continuous refinement of AI models and systems.
- Few-shot PromptingA prompting technique that guides a language model by including a small number of input-output examples in the prompt, helping the model infer the desired task, format, or behavior.
- Files APIAn API that enables applications to upload, store, retrieve, manage, and reference files for use in AI workflows such as fine-tuning, retrieval, batch processing, and multimodal inference.
- Fine-tuningThe process of further training a pretrained AI model on a task-specific or domain-specific dataset to improve its performance, behavior, or specialization for particular use cases.
- Fine-tuning APIAn API that enables developers to create, manage, and deploy customized AI models by training a pretrained model on task-specific datasets.
- Fireworks AIAn AI infrastructure and model platform that provides high-performance inference, model hosting, fine-tuning, and deployment services for open-source and proprietary foundation models.
- Fireworks AI SDKA software development kit provided by Fireworks AI that enables developers to integrate, manage, and interact with Fireworks AI models and inference services through programmatic APIs.
- FlowiseAn open-source low-code framework for building, orchestrating, and deploying AI applications and agent workflows using a visual drag-and-drop interface with support for large language models, tools, memory, and vector databases.
- Function CallingA capability that enables an AI model to identify when an external function or tool should be invoked, generate the required structured arguments, and incorporate the returned results into its workflow or response.
- Function SchemaA structured specification that defines a callable function's name, purpose, parameters, input types, constraints, and expected outputs, enabling AI systems to correctly invoke tools and generate valid arguments.
- Functional TestingA software testing method that verifies whether an AI application, agent, or system performs its intended functions correctly by validating its behavior against specified functional requirements and expected outcomes.
G
- GalileoAn AI observability and evaluation platform that helps developers monitor, evaluate, debug, and improve the quality, reliability, and performance of large language model (LLM) applications and AI systems.
- GatewayAn interoperability pattern that acts as a centralized entry point between systems, translating, routing, securing, or managing requests to simplify communication and integration across multiple services or APIs.
- GitOpsA software delivery practice that uses Git repositories as the single source of truth for declarative infrastructure and application configurations, enabling automated deployment, version control, and continuous reconciliation through Git-based workflows.
- Goal AutonomyThe capability of an AI agent to independently determine, prioritize, or adapt its goals within defined objectives, constraints, or policies without requiring continuous human direction.
- Goal DecompositionThe planning process of breaking a high-level goal into smaller, manageable subgoals or tasks that can be executed sequentially, in parallel, or hierarchically to achieve the overall objective.
- Goal PlanningThe process by which an AI system identifies, prioritizes, and organizes objectives into an actionable plan, determining the sequence of tasks and strategies required to achieve one or more specified goals.
- Golden DatasetA curated and validated collection of high-quality reference examples with trusted labels or expected outputs that serves as a benchmark for evaluating, testing, and monitoring the performance of AI models and applications.
- Google DeepMindAn AI research and technology company that develops foundation models, multimodal AI systems, and machine learning technologies, providing models and AI services for research, enterprise, and developer applications.
- Google Gen AI SDKA software development kit provided by Google for building AI applications using Gemini models and related AI services, offering APIs for text, image, multimodal, embedding, and streaming capabilities across supported programming languages.
- Governance PolicyA formal set of rules, principles, and operational requirements that defines how an AI system is developed, deployed, monitored, and controlled to ensure compliance, safety, accountability, and responsible operation.
- GPQAGPQAA benchmark that evaluates the ability of AI models to answer challenging graduate-level multiple-choice questions across scientific domains, measuring expert-level reasoning, scientific knowledge, and problem-solving performance.
- GPUGPUA Graphics Processing Unit (GPU) is a specialized parallel processor designed to perform large-scale mathematical computations efficiently, making it the primary hardware for training and running AI models.
- GPU DeploymentThe practice of deploying AI models or applications on graphics processing units (GPUs) to accelerate inference or training by leveraging massively parallel computation for high-performance workloads.
- Graph of ThoughtsGoTA reasoning strategy in which an AI system organizes thoughts as a graph of interconnected nodes rather than a single sequence, allowing it to explore, combine, and revisit multiple reasoning paths before reaching a conclusion.
- GraphQLA query language and API protocol that enables clients to request exactly the data they need from a server through a single endpoint, providing flexible, efficient, and strongly typed communication between applications and services.
- Grid SearchA hyperparameter optimization technique that systematically evaluates all predefined combinations of parameter values to identify the configuration that produces the best performance for an AI model or system.
- GroqAn AI infrastructure company and model platform that provides ultra-low-latency inference services for large language models and other AI models, optimized through its custom Language Processing Unit (LPU) architecture.
- Groq SDKA software development kit provided by Groq that enables developers to integrate, manage, and interact with Groq's AI models and high-speed inference services through programmatic APIs.
- GroundednessAn evaluation metric that measures whether an AI system's output is supported by the provided context, retrieved information, or source material without introducing unsupported claims or hallucinations.
- Grounding VerificationThe process of verifying that an AI system's responses are supported by the provided context, retrieved knowledge, or authoritative sources before they are presented or acted upon, helping prevent hallucinations and unsupported claims.
- gRPCgRPCA high-performance remote procedure call (RPC) framework that uses Protocol Buffers for interface definition and data serialization, enabling efficient, strongly typed communication between distributed applications and services.
- GSM8KGSM8KA benchmark that evaluates the ability of AI models to solve grade school mathematical word problems requiring multi-step reasoning, arithmetic, and logical problem-solving.
- GuidanceAn open-source Python library for controlling and constraining large language model outputs using structured prompts, grammars, templates, and programmatic generation, enabling more reliable and deterministic AI applications.
H
- Hallucination DetectionThe process of identifying AI-generated outputs that contain fabricated, inaccurate, or unsupported information by comparing responses against trusted sources, retrieved context, or established facts before or after they are presented.
- Hallucination RateAn evaluation metric that measures the proportion of an AI system's outputs that contain fabricated, unsupported, or factually incorrect information relative to a trusted reference, provided context, or authoritative source.
- Hallucination TestingA testing process that evaluates an AI system's tendency to generate fabricated, unsupported, or factually incorrect information by comparing its outputs against trusted references, provided context, or authoritative sources.
- HaystackAn open-source AI framework for building production-ready applications powered by large language models, retrieval-augmented generation (RAG), semantic search, and intelligent agent pipelines using modular components for retrieval, prompting, and orchestration.
- Health CheckA mechanism that periodically verifies whether a service, application, model endpoint, or AI agent is operational, responsive, and able to perform its intended functions, enabling orchestration systems to detect failures and route traffic appropriately.
- HeartbeatA periodic signal sent by a service, AI agent, or system component to indicate that it is active, healthy, and reachable, enabling orchestration and monitoring systems to detect failures, timeouts, or disconnected instances.
- HeliconeAn AI observability platform that provides logging, monitoring, analytics, caching, cost tracking, prompt management, and evaluation capabilities for applications built with large language models.
- HellaSwagA benchmark that evaluates the commonsense reasoning and natural language understanding capabilities of AI models by requiring them to select the most plausible continuation of a given context from multiple candidate endings.
- HelmAn open-source package manager for Kubernetes that simplifies the deployment, configuration, upgrading, and management of containerized applications using reusable packages called Helm Charts.
- Helm ChartA packaged collection of Kubernetes resource templates, configuration values, and metadata that defines how an application or AI service is installed, configured, upgraded, and managed using Helm.
- HelpfulnessAn evaluation metric that measures how effectively an AI system's response addresses the user's request by providing relevant, accurate, complete, and actionable information that satisfies the intended task or objective.
- Hierarchical AgentAn AI agent architecture that organizes multiple agents into hierarchical layers, where higher-level agents decompose goals, coordinate decision-making, and supervise lower-level agents responsible for executing specialized tasks.
- Hierarchical PlanningA planning approach in which a complex goal is recursively decomposed into smaller subgoals and tasks, allowing an AI system to create and execute plans at multiple levels of abstraction.
- HoneyHiveAn AI observability and evaluation platform that helps developers monitor, evaluate, test, and optimize large language model applications through tracing, experiment management, prompt evaluation, and performance analytics.
- HTTPHTTPThe Hypertext Transfer Protocol (HTTP) is an application-layer communication protocol used to exchange requests and responses between clients and servers over a network, forming the foundation of web APIs, websites, and distributed applications.
- HTTP/2HTTP/2The second major version of the Hypertext Transfer Protocol (HTTP) that improves web communication performance through features such as multiplexing, header compression, and binary framing, enabling more efficient communication between clients and servers.
- HTTP/3HTTP/3The third major version of the Hypertext Transfer Protocol (HTTP) that uses the QUIC transport protocol over UDP to provide lower latency, faster connection establishment, improved reliability, and better performance for modern web applications and distributed services.
- HTTPSHTTPSThe Hypertext Transfer Protocol Secure (HTTPS) is the secure version of HTTP that encrypts communication between clients and servers using Transport Layer Security (TLS), ensuring data confidentiality, integrity, and authentication during transmission.
- Hub-and-Spoke ArchitectureAn AI system architecture in which a central coordinator, or hub, manages communication, task delegation, and orchestration among multiple specialized agents, or spokes, that independently execute domain-specific tasks and report results back to the hub.
- Hugging FaceAn AI platform and company that develops and hosts open-source machine learning models, datasets, and developer tools, providing infrastructure and services for building, deploying, sharing, and running AI applications.
- Hugging Face HubA platform and API provided by Hugging Face for discovering, sharing, versioning, and accessing machine learning models, datasets, Spaces, and related AI resources, with SDKs that enable seamless integration into AI applications.
- Human Approval WorkflowA workflow in which one or more human reviewers must approve, modify, or reject an AI system's decisions or actions before execution, ensuring oversight, quality control, compliance, or risk mitigation.
- Human EvaluationAn evaluation method in which human reviewers assess the quality of AI system outputs against defined criteria such as correctness, relevance, helpfulness, safety, or fluency, providing judgments that complement or validate automated evaluation metrics.
- Human FeedbackInformation, evaluations, corrections, or preferences provided by human users or reviewers about an AI system's outputs or behavior, which can be used to improve model performance, refine prompts, optimize workflows, or guide future decisions.
- Human OversightThe practice of monitoring, reviewing, and governing an AI system's decisions, actions, and outputs by human operators to ensure safety, accuracy, compliance, accountability, and appropriate intervention when necessary.
- Human-in-the-LoopHITLA collaborative approach in which humans actively participate in an AI system's operation by reviewing, approving, correcting, or guiding its decisions or actions, improving accuracy, safety, accountability, and overall system performance.
- Human-on-the-LoopHOTLA collaborative approach in which an AI system operates autonomously while humans supervise its behavior, monitor outcomes, and intervene only when necessary to correct, override, or stop its decisions or actions.
- HumanEvalA benchmark that evaluates the code generation capabilities of AI models by measuring their ability to generate functionally correct code that passes predefined unit tests for a collection of programming tasks.
- HumanloopAn AI development and observability platform that enables teams to build, evaluate, monitor, and improve large language model applications through prompt management, human evaluation, experimentation, tracing, and production analytics.
- Hybrid SearchA retrieval technique that combines semantic vector search with keyword-based lexical search to improve the relevance, accuracy, and recall of retrieved information by leveraging the strengths of both approaches.
- Hyperparameter OptimizationHPOThe process of automatically or systematically searching for the combination of hyperparameter values that produces the best performance for an AI model or system according to one or more evaluation metrics.
- Hyperparameter TuningThe process of selecting and adjusting the hyperparameters of an AI model or system through controlled experiments to identify the configuration that delivers the best performance according to one or more evaluation metrics.
I
- Identity ProviderIdPA system or service that authenticates users or applications and provides verified identity information to other systems, enabling secure access through protocols such as OAuth, OpenID Connect (OIDC), or SAML.
- Image RegistryA repository service for storing, managing, versioning, and distributing container images, enabling consistent deployment of AI applications and services across development, testing, and production environments.
- Impact AssessmentA structured process for evaluating the potential effects, risks, and consequences of an AI system on individuals, organizations, society, or the environment, supporting informed decisions about deployment, governance, compliance, and risk mitigation.
- Implicit FeedbackFeedback inferred from user behavior or interactions rather than explicitly provided ratings or comments, such as clicks, dwell time, task completion, corrections, or repeated usage, which can be used to evaluate and improve AI systems.
- Indexing PipelineA sequence of processes that ingests, cleans, transforms, chunks, enriches, embeds, and indexes data into a searchable store, preparing documents and other content for efficient retrieval by AI applications.
- Inductive ReasoningIRA reasoning approach in which an AI system infers general rules, patterns, or hypotheses from specific observations, examples, or experiences, producing conclusions that are probable rather than logically guaranteed.
- Inference APIAn API that enables applications to send input data to a deployed AI model and receive generated predictions or outputs, providing programmatic access to inference capabilities without managing the underlying model infrastructure.
- Inference OptimizationThe process of improving the speed, latency, throughput, memory usage, or cost-efficiency of AI model inference through techniques such as quantization, pruning, batching, caching, hardware acceleration, and optimized model serving.
- Inference ServerA software service that hosts one or more AI models and processes inference requests by receiving inputs, executing model predictions, and returning outputs through APIs while managing resources, batching, scaling, and hardware acceleration.
- Infrastructure as CodeIaCA software engineering practice in which infrastructure resources such as servers, networks, cloud services, and Kubernetes clusters are defined, provisioned, and managed using declarative or imperative code instead of manual configuration.
- Input ValidationThe process of verifying and sanitizing user inputs or external data before they are processed by an AI system to ensure they conform to expected formats, constraints, and security policies, helping prevent errors, misuse, prompt injection, and malicious inputs.
- Instruction PromptingA prompting technique in which clear, explicit instructions are provided to guide an AI model toward performing a specific task, following defined constraints, and producing outputs in the desired format or style.
- InstructorAn open-source Python library for structured output generation that enables developers to reliably extract validated data from large language models by defining schemas with Pydantic and automatically parsing model responses into typed objects.
- Integration AdapterA software component that translates data formats, protocols, or interfaces between an AI application and an external system, enabling interoperability without requiring changes to either side of the integration.
- Integration LayerAn architectural layer that mediates communication between applications, services, AI agents, and external systems by handling data transformation, protocol translation, routing, authentication, and orchestration to enable seamless interoperability.
- Integration TestingA testing method that verifies multiple software components, services, APIs, AI models, or external systems work together correctly by validating their interactions, data flow, and end-to-end behavior after individual components have been tested.
- Interactive PlanningA collaborative planning approach in which humans and AI systems iteratively develop, refine, and adjust plans together through ongoing feedback, allowing users to guide objectives, constraints, priorities, and execution decisions.
- Interface ContractA formal specification that defines how software components, services, APIs, or AI agents interact by describing supported operations, input and output formats, data types, protocols, constraints, and expected behavior, ensuring consistent and interoperable communication.
- InterpretabilityThe ability to understand how an AI model or system arrives at its predictions, decisions, or outputs by examining its internal mechanisms, decision-making process, or contributing factors.
J
- Jailbreak DetectionThe process of identifying prompts or interactions that attempt to bypass an AI system's safety policies, behavioral constraints, or security controls, enabling the system to block, flag, or mitigate unauthorized or unsafe requests.
- Jailbreak TestingA security testing method that evaluates whether an AI system can be manipulated into bypassing its safety policies, behavioral constraints, or security guardrails through adversarial prompts or other attack techniques.
- JAXJAXAn open-source Python library developed by Google for high-performance numerical computing and machine learning, providing automatic differentiation, just-in-time (JIT) compilation, vectorization, and execution on CPUs, GPUs, and TPUs.
- JSONJSONJavaScript Object Notation (JSON) is a lightweight, text-based data interchange format for representing structured data using key-value pairs and arrays. It is widely used for exchanging data between applications, APIs, AI systems, and web services.
- JSON SchemaA declarative specification for describing the structure, data types, constraints, and validation rules of JSON documents, enabling AI systems, APIs, and tools to exchange structured data reliably and validate inputs and outputs.
- JSON Web TokenJWTA compact, URL-safe token format used to securely transmit signed or encrypted claims between parties, commonly for authentication, authorization, and identity propagation in web APIs and distributed systems.
- JSON-RPCJSON-RPCA lightweight remote procedure call (RPC) protocol that uses JSON to encode requests and responses, enabling clients to invoke methods on remote servers over transports such as HTTP, WebSocket, or TCP.
- JsonformerAn open-source Python library for constrained text generation that enables large language models to produce JSON outputs conforming to a predefined JSON Schema by generating only the content tokens while enforcing the required structure.
K
- Key RotationThe security practice of periodically replacing cryptographic keys, API keys, certificates, or other credentials with new ones to reduce the risk of compromise, limit the impact of credential exposure, and maintain secure access to systems and services.
- Kill SwitchA safety mechanism that allows an AI system, service, or automated workflow to be immediately disabled or halted in response to unsafe behavior, security incidents, operational failures, or emergency conditions, preventing further execution until the issue is resolved.
- Knowledge BaseA structured repository of documents, facts, policies, procedures, or other information that an AI system can retrieve and use to answer questions, generate responses, or support decision-making with relevant and up-to-date knowledge.
- Knowledge DistillationA model optimization technique in which a smaller AI model, called the student model, is trained to mimic the behavior or outputs of a larger, more capable teacher model, reducing model size and inference cost while preserving much of the original performance.
- KubernetesK8sAn open-source container orchestration platform that automates the deployment, scaling, networking, and management of containerized applications across clusters of machines, making it a standard platform for running production AI and cloud-native workloads.
- Kubernetes ClusterA group of interconnected machines running Kubernetes that work together to deploy, schedule, scale, and manage containerized applications across multiple worker nodes under the control of a centralized control plane.
- Kubernetes ManifestA declarative configuration file, typically written in YAML, that defines the desired state of Kubernetes resources such as Pods, Deployments, Services, ConfigMaps, and Ingresses, enabling consistent deployment and management of containerized applications.
- KustomizeAn open-source Kubernetes configuration management tool that enables users to customize declarative Kubernetes manifests through overlays, patches, and reusable configuration without modifying the original resource definitions.
L
- LabelingThe process of assigning structured annotations, categories, ratings, or ground-truth values to data or AI outputs so they can be used for training, evaluation, validation, or continuous improvement of AI systems.
- LanceDBAn open-source vector database designed for AI applications that stores vector embeddings alongside structured data, enabling efficient similarity search, hybrid search, indexing, and retrieval for retrieval-augmented generation (RAG) and semantic search workloads.
- LangChainAn open-source framework for building applications powered by large language models, providing components for prompt management, model integration, tool calling, agents, memory, retrieval, and workflow orchestration across multiple AI providers.
- LangfuseAn observability platform for tracing, monitoring, evaluating, and debugging LLM applications and agentic systems.
- LangGraphA framework for building stateful, multi-step AI agent workflows as graphs of nodes and edges.
- LangSmithA platform for tracing, evaluating, monitoring, and debugging LLM and agent applications.
- LatencyThe time delay between an AI system receiving a request and producing a response or completing an operation.
- Latency MonitoringThe practice of tracking response latency over time to detect slowdowns, performance degradation, and service issues.
- Latency OptimizationThe practice of reducing response latency in AI systems to improve responsiveness, throughput, and user experience.
- Lexical SearchA search method that retrieves documents by matching query terms against exact words, tokens, or linguistic forms in indexed content.
- LiteLLMA library that provides a unified interface for interacting with multiple large language model providers and APIs.
- LiveCodeBenchA contamination-aware benchmark for evaluating large language models on recent competitive programming problems.
- llama.cppA lightweight C/C++ library for running and deploying large language models locally across a range of hardware.
- LlamaIndexA framework for building LLM applications that connect language models with external data, knowledge sources, and retrieval workflows.
- LLM-as-a-JudgeAn evaluation method that uses a large language model to assess the quality, correctness, or other attributes of AI-generated outputs.
- LM StudioA desktop application for discovering, downloading, and running large language models locally on consumer hardware.
- Load BalancerA system component that distributes incoming requests across multiple service instances to improve availability, scalability, and resource utilization.
- Load BalancingThe practice of distributing workloads or requests across multiple computing resources to improve scalability, availability, and resource utilization.
- Load TestingA testing method that evaluates system behavior and performance under expected levels of concurrent requests or workload.
- Local DevelopmentThe practice of building, testing, and running software on a developer's local machine before deployment to shared environments.
- Local EmulatorA local tool that simulates an external service or runtime environment for development and testing without requiring the actual remote system.
- LoggingThe practice of recording application, model, and system events to support debugging, monitoring, troubleshooting, and operational analysis.
- Long PollingA communication technique where a client holds a request open until the server has new data or a timeout occurs, reducing repeated polling requests.
- Long-running WorkflowA workflow designed to execute tasks that may continue for extended periods, often across asynchronous steps, interruptions, or persistent state.
- Long-term MemoryLTMA memory system that stores information over extended periods, enabling AI systems to retain knowledge and retrieve it across interactions.
- LongBenchA benchmark for evaluating large language models on tasks requiring understanding and reasoning over long contexts.
- LoRALoRAA parameter-efficient fine-tuning method that adapts a model by training low-rank matrices while keeping the original model weights frozen.
M
- MastraA TypeScript framework for building AI applications and agent workflows with tools, memory, workflows, and integrations.
- MATHMATHA benchmark for evaluating mathematical problem-solving abilities of AI models across challenging competition-level mathematics problems.
- MBPPMBPPA benchmark for evaluating the ability of language models to generate Python code from natural language descriptions.
- MCP ClientMCPA software component that connects to an MCP server to discover and use its exposed tools, resources, and prompts.
- MCP HostMCPA host application that manages MCP clients and provides AI models with access to tools, resources, and prompts exposed by MCP servers.
- MCP PromptMCPA reusable prompt template exposed by an MCP server that applications can discover and use to guide interactions with AI models.
- MCP ResourceMCPA data source exposed by an MCP server that provides structured context or information for use by connected AI applications.
- MCP ServerMCPA server that implements the Model Context Protocol to expose tools, resources, and prompts to connected AI applications.
- MCP ToolMCPA callable capability exposed by an MCP server that allows an AI application to perform an action or retrieve information.
- MediatorAn integration pattern that centralizes communication between components to coordinate interactions without requiring direct connections between them.
- Memory ConsolidationThe process of stabilizing and integrating information into longer-term memory for retention and later retrieval.
- Memory RetrievalThe process by which an AI system accesses relevant information from stored memory to support context-aware responses and actions.
- Memory StoreA system or component where an AI system retains information for future retrieval and use.
- Message BrokerA system that receives, routes, and delivers messages between producers and consumers, decoupling services that communicate asynchronously.
- Message PassingA communication mechanism that enables agents or software components to exchange messages, information, and task-related data.
- Message QueueA communication mechanism that stores messages temporarily so producers and consumers can exchange data asynchronously without direct coordination.
- MessagePackA compact binary serialization format for exchanging structured data efficiently between applications and services.
- Meta AIAn AI research and product organization at Meta that develops foundation models, generative AI systems, and AI-powered products.
- Metadata FilteringA retrieval technique that filters documents or chunks using metadata attributes before or during semantic or lexical search.
- METEORMETEORA metric for evaluating generated text by measuring alignment with reference text using word matches, stemming, and synonym matching.
- Metrics CollectionThe process of gathering quantitative measurements from AI systems and applications to track performance, reliability, usage, and operational health.
- MilvusAn open-source vector database designed to store, index, and search high-dimensional vector embeddings for AI applications.
- MIME TypeMIMEA standardized identifier that specifies the format and nature of data exchanged between applications, especially in network requests and responses.
- Mistral AIAn AI company that develops and provides open and commercial large language models and generative AI systems.
- Mistral SDKA software development kit for integrating Mistral AI models and services into applications through programmatic APIs.
- MLflowAn open-source platform for tracking, evaluating, managing, and deploying machine learning and AI models and workflows.
- MMBenchMMBenchA benchmark for evaluating multimodal large language models across a broad range of vision-language tasks and capabilities.
- MMLUMMLUA benchmark for evaluating language models across diverse subjects spanning humanities, social sciences, STEM, and professional knowledge.
- MMLU-ProA more challenging version of MMLU designed to evaluate advanced reasoning and knowledge across diverse academic and professional subjects.
- MMMUMMMUA multimodal benchmark for evaluating models on complex tasks spanning diverse academic disciplines and requiring both visual and textual reasoning.
- MMMU-ProA challenging multimodal benchmark designed to evaluate advanced reasoning across diverse academic and professional tasks.
- Mock ServerA simulated server that mimics an API or service for development and testing without relying on the actual backend.
- MockingA development technique that replaces real dependencies with controlled simulated components for isolated software development and testing.
- Model CardA document that describes an AI model's intended uses, limitations, capabilities, evaluation results, and other relevant information.
- Model ComparisonAn evaluation method that compares AI models on the same tasks, datasets, or metrics to identify differences in capabilities and performance.
- Model Context ProtocolMCPAn open protocol that standardizes how AI applications connect to external tools, resources, and prompts.
- Model MonitoringThe practice of continuously tracking an AI model's performance, behavior, usage, and operational health in production.
- Model OptimizationThe process of improving an AI model's performance, efficiency, or resource usage through techniques such as fine-tuning, pruning, quantization, or distillation.
- Model RegistryA centralized system for storing, versioning, organizing, and managing AI models and their deployment metadata.
- Model ServingThe process of making an AI model available for inference by deploying it behind an interface that accepts requests and returns predictions or generated outputs.
- Model-based EvaluationAn evaluation approach that uses another trained model to assess the quality, correctness, safety, or other properties of an AI system's outputs.
- ModerationThe process of detecting, classifying, and handling harmful, unsafe, or policy-violating content in AI inputs and outputs.
- MongoDB Atlas Vector SearchA vector search capability in MongoDB Atlas for storing, indexing, and querying vector embeddings alongside application data.
- MQTTMQTTA lightweight publish-subscribe messaging protocol designed for efficient communication between constrained devices and unreliable networks.
- MT-BenchMT-BenchA benchmark for evaluating conversational AI models through multi-turn dialogue tasks that assess instruction following and response quality.
- Multi-Agent SystemMASAn AI architecture in which multiple autonomous agents interact and collaborate to achieve individual or shared goals.
- Multivariate TestingAn experimentation method that tests multiple variables and their combinations simultaneously to measure their effects on an AI system or product.
- Mutual TLSmTLSA TLS authentication method in which both client and server verify each other's identities using digital certificates before establishing a secure connection.
N
- NDJSONNDJSONA line-delimited format where each line contains a separate JSON object, enabling efficient streaming and incremental processing of structured data.
- NegotiationA coordination process in which agents exchange proposals and preferences to reach an acceptable agreement or resolve competing goals.
- Network StorageStorage that provides data access over a network, allowing applications or services to share persistent files or volumes across systems.
- Notebook EnvironmentAn interactive development environment that combines executable code, documentation, outputs, and visualizations in a notebook-based workflow.
- NumPyA Python library for efficient numerical computing with multidimensional arrays, mathematical operations, and scientific data processing.
O
- OAuthAn authorization framework that enables applications to obtain delegated access to protected resources without exposing user credentials.
- OAuth 2.0OAuthAn authorization framework that enables applications to obtain limited access to protected resources without exposing a user's credentials.
- Object StorageA storage system that manages data as objects with associated metadata, providing scalable and durable access over networks.
- OCI ImageOCIA container image packaged according to the Open Container Initiative specification for portable distribution and execution across compatible runtimes.
- Offline EvaluationAn evaluation performed outside production using fixed datasets, test cases, or recorded interactions to assess an AI system before deployment.
- OllamaA tool for running and managing large language models locally, providing a simple interface for downloading, serving, and interacting with models.
- On-device AIAI capabilities that run directly on local devices, reducing reliance on remote servers and enabling lower-latency, privacy-focused inference.
- One-shot PromptingA prompting technique that provides a single example demonstrating the desired task or output format before asking a model to perform the task.
- Online EvaluationAn evaluation performed on live or production-like interactions to continuously assess an AI system's behavior, quality, safety, or performance.
- Online LearningA learning approach in which an AI system continuously updates its behavior or model using newly available data during operation.
- OpenAIAn AI research and deployment company that develops foundation models, generative AI systems, and tools for building AI applications.
- OpenAI SDKSDKA software development kit for integrating OpenAI models and APIs into applications through supported programming languages and client libraries.
- OpenAPIOASA machine-readable specification for describing HTTP APIs, enabling AI systems and applications to discover and integrate callable API operations.
- OpenAPI SpecificationOASA standard for describing HTTP APIs in a machine-readable format, enabling documentation, client generation, testing, and integration.
- OpenCVOpenCVAn open-source computer vision library for image processing, video analysis, object detection, and other visual computing tasks.
- OpenDevinAn open-source framework for building AI software engineering agents that perform coding and development tasks.
- OpenHandsAn open-source framework for building AI software engineering agents that can perform coding and development tasks.
- OpenID ConnectOIDCAn identity layer built on OAuth 2.0 that enables applications to verify user identity and obtain basic profile information.
- OpenLITAn open-source observability platform for monitoring and analyzing AI applications, including model calls, traces, metrics, and performance.
- OpenRouterAn AI model routing platform that provides a unified API for accessing and switching between models from multiple providers.
- OpenSearch Vector EngineA vector search engine in OpenSearch for storing, indexing, and searching high-dimensional vector embeddings for similarity-based retrieval.
- OpenTelemetryOTelAn open standard for collecting, exporting, and correlating telemetry data such as traces, metrics, and logs across distributed systems.
- Operational AutonomyThe ability of an AI system to independently perform operational tasks and actions within defined goals, permissions, and constraints.
- OrchestratorA system component that coordinates and manages services, workflows, or tasks to control how distributed operations are executed.
- Orchestrator AgentAn AI agent designed to coordinate and manage other agents to achieve complex goals.
- OutlinesA Python library for guiding language models to generate structured text that follows specified formats, schemas, or constraints.
- Output ValidationThe process of checking AI-generated outputs against expected formats, schemas, rules, or safety requirements before they are accepted or used.
P
- Package ManagerA tool that automates installing, updating, removing, and managing software packages and their dependencies in a development environment.
- Package RegistryA service that stores and distributes software packages, allowing developers and package managers to publish, discover, and install dependencies.
- Pair ProgrammingA collaborative development practice where two participants work together on the same code, with one writing code while the other reviews and guides.
- Pairwise ComparisonAn evaluation method that compares two outputs directly to determine which performs better against a defined criterion or preference.
- PandasAn open-source Python library for data manipulation and analysis, providing data structures and tools for working with structured datasets.
- Parallel WorkflowA workflow that executes multiple independent tasks or steps concurrently to reduce overall execution time.
- Parameter SweepAn experiment that evaluates a system across a defined range of parameter values to compare results and identify effective configurations.
- Pass@kA code-generation metric measuring the probability that at least one of k generated solutions correctly solves a given programming problem.
- PayloadThe data carried within a request, response, message, or event exchanged between systems.
- Peer-to-Peer AgentsAgents that communicate and collaborate directly with one another without requiring a central coordinating agent.
- PEFTPEFTA library for parameter-efficient fine-tuning of pretrained models by updating a small subset of parameters instead of the full model.
- Performance TestingTesting that measures how an AI system performs under defined workloads, including response time, throughput, resource usage, and scalability.
- Performance TuningThe process of improving an AI system's speed, resource efficiency, throughput, or responsiveness by adjusting its implementation and runtime configuration.
- Permission BoundaryA defined limit on the actions, resources, or capabilities an AI agent is permitted to access or execute.
- Permission ScopeA defined set of permissions that limits the resources or actions an authenticated identity or client is authorized to access.
- PerplexityA metric that measures how well a language model predicts a sequence of tokens, with lower values indicating better predictive performance.
- Persistent ContextContext that is retained across sessions or interactions so an AI system can reuse relevant information over time.
- Personal AgentAn AI agent designed to assist an individual user by managing personal tasks, preferences, and information.
- pgvectorAn open-source PostgreSQL extension that adds vector data types and similarity search capabilities for storing and retrieving embeddings.
- PII DetectionPIIA guardrail that identifies personally identifiable information in inputs or outputs so it can be protected, redacted, or blocked.
- PineconeA managed vector database designed to store, index, and search high-dimensional embeddings for semantic retrieval and AI applications.
- Plan ExecutionThe process of carrying out the actions and steps defined by an AI agent's plan to achieve a specified goal.
- Planner-Executor ArchitectureAn AI architecture that separates planning from execution, with one component generating a plan and another carrying out its actions.
- Planning AgentAn AI agent designed to create plans and strategies for achieving specific goals.
- PluginA software component that extends an application or AI system with additional functionality through a defined integration interface.
- Plugin InterfaceA defined interface that allows plugins to extend an AI system with additional capabilities while conforming to expected integration contracts.
- PocketFlowA lightweight framework for building AI workflows by composing modular nodes and flows into structured execution pipelines.
- Pointwise EvaluationAn evaluation method that assigns an independent score to each individual model response using defined criteria, without comparing it directly to another response.
- Policy EnforcementThe process of applying defined rules and policies to control, restrict, or permit AI system inputs, outputs, and actions.
- Port and AdapterAn architectural pattern that isolates core application logic from external systems through ports and interchangeable adapters.
- Power AnalysisA statistical method for determining whether an experiment has enough observations to reliably detect an effect of a specified size.
- PrecisionA classification metric measuring the proportion of predicted positive cases that are actually positive.
- Preference DataData that records preferences between AI outputs or behaviors, typically used to train or optimize models toward preferred responses.
- Preference LearningA learning approach in which an AI system learns from preferences between possible outputs or actions to better align its behavior with desired outcomes.
- Preference ModelingA method for representing and estimating preferences between possible actions or outcomes to guide an AI system's decision-making.
- Preference OptimizationThe process of optimizing an AI system to produce outputs that better match preferred responses, behaviors, or outcomes.
- PrivacyThe protection and appropriate handling of personal or sensitive information throughout an AI system's collection, processing, storage, and use.
- Procedural MemoryPMA form of memory that stores knowledge about how to perform tasks or actions, enabling an AI system to reuse learned procedures.
- PromptAn input containing instructions, context, or data provided to an AI model to guide its response or behavior.
- Prompt CachingA technique that stores reusable prompt content so repeated requests can avoid processing the same input repeatedly, reducing latency and inference cost.
- Prompt ChainingA technique that links multiple prompts so the output of one prompt becomes the input to the next, enabling complex tasks through sequential steps.
- Prompt CompressionA technique for reducing the size of a prompt while preserving the information needed for an AI model to produce the desired output.
- Prompt ExperimentA structured experiment that compares prompt variants to measure their effects on AI system behavior, quality, or performance.
- Prompt InjectionAn attack that inserts malicious or unintended instructions into prompts or external content to manipulate an AI system's behavior.
- Prompt Injection DetectionA guardrail that detects attempts to manipulate an AI system through malicious or unintended instructions embedded in prompts or external content.
- Prompt LibraryA curated collection of reusable prompts organized for discovery, reuse, versioning, and consistent AI application development.
- Prompt OptimizationThe process of systematically improving prompts to increase the quality, consistency, efficiency, or reliability of AI model outputs.
- Prompt PlaygroundAn interactive development environment for creating, testing, and comparing prompts while observing AI model outputs.
- Prompt ShieldA guardrail that detects or blocks malicious prompt content and instruction-based attacks before they can influence an AI system.
- Prompt TemplateA reusable prompt structure containing fixed instructions and variable placeholders for generating consistent inputs to an AI model.
- Prompt TestingA testing practice that evaluates prompts across defined inputs and criteria to assess the quality, consistency, and reliability of AI model outputs.
- Prompt VersioningThe practice of tracking and managing different versions of prompts to preserve changes, compare variants, and maintain reproducibility.
- PromptLayerA developer tool for tracking, managing, and observing prompts and LLM interactions across AI applications.
- Protocol BuffersProtobufA language- and platform-neutral mechanism developed by Google for serializing structured data.
- Protocol TranslationThe process of converting messages and data formats between different communication protocols to allow disparate systems to interact.
- ProxyAn intermediary service that receives requests, forwards them to appropriate target systems, and returns the response.
- PruningThe process of removing unnecessary parameters or weights from a model to reduce size and speed up inference.
- Publish-SubscribePub/SubA messaging pattern where senders publish messages to topics without targeting specific receivers, which subscribe to receive them.
- PydanticA data validation and settings management library for Python using type annotations.
- PydanticAIA Python agent framework designed by Pydantic for building production-grade generative AI applications.
- PyTorchAn open-source machine learning framework designed for deep learning, tensor computation, and dynamic neural network construction.
Q
- QdrantAn open-source vector similarity search engine and database designed for managing high-dimensional vector embeddings.
- QuantizationThe process of reducing the precision of a model's weights and activations to lower memory footprint and speed up inference.
- Query ExpansionThe technique of augmenting or reformulating an input search query with related terms, synonyms, or context to improve retrieval recall.
- Query RewritingThe process of transforming or optimizing a user search prompt to improve the retrieval of relevant context and documents.
- Queue ManagerA system component that controls the storage, ordering, and delivery of asynchronous messages and tasks.
- QUICQUICA transport layer network protocol designed to reduce latency and improve connection performance over UDP.
R
- Random SearchAn optimization technique that evaluates randomly selected hyperparameter configurations from a defined search space.
- Rapid PrototypingAn iterative workflow focused on quickly building minimal functional versions of software to test, validate, and refine concepts.
- Rate LimiterA system component that controls the rate of incoming or outgoing traffic to prevent service overload and resource exhaustion.
- Reactive AgentAn AI agent that responds directly to immediate environmental stimuli in real time without relying on internal state models or long-term planning.
- Realtime APIAn API enabling bidirectional, low-latency streaming of audio and text data for real-time interactive applications.
- RecallA performance metric that measures the proportion of actual positive instances correctly identified by a model.
- Red TeamingA structured testing methodology where adversarial tactics are simulated to discover vulnerabilities, safety flaws, and failure modes in a system.
- Redis Vector SearchAn extension to Redis that provides high-performance vector indexing and similarity search for real-time applications.
- Reference-based EvaluationAn evaluation method that measures model output quality by comparing generated responses against ground-truth reference data or golden datasets.
- Reference-free EvaluationAn evaluation approach that assesses model output quality without relying on ground-truth reference answers or golden datasets.
- ReflectionA reasoning process in which an AI system evaluates its own thoughts and outputs to refine decisions and improve performance.
- Refresh TokenA long-lived credential used to request new access tokens without requiring the user to re-authenticate.
- Registry FederationThe linking of multiple decentralized service or agent registries to enable discovery and query resolution across domains.
- Registry LookupThe action or process of querying an agent or service registry to find and retrieve metadata or connection endpoints.
- Registry SynchronizationThe process of updating and maintaining consistent agent or service capability metadata across distributed registries.
- Regression EvaluationAn evaluation technique that checks whether system updates or model iterations introduce unexpected performance degrades.
- Regression TestingA software testing practice that verifies recent code changes have not negatively affected existing functionality or introduced new bugs.
- Regulatory ComplianceThe practice of ensuring AI systems comply with legal mandates, industry regulations, and standards.
- Reinforcement LearningRLA machine learning paradigm where an agent learns to make optimal decisions by taking actions in an environment to maximize cumulative rewards.
- Reinforcement Learning from AI FeedbackRLAIFA fine-tuning method that optimizes model behavior using feedback generated by another AI system rather than human annotators.
- Reinforcement Learning from Human FeedbackRLHFA machine learning alignment technique that optimizes model behavior based on preferences gathered from human evaluators.
- RelevanceA performance metric that measures how pertinent and applicable generated outputs or retrieved results are to the query.
- ReliabilityA performance metric measuring the degree to which a system consistently performs its intended function without failure over time.
- REPLREPLAn interactive programming environment that takes user inputs, executes them, and returns the result to the user.
- ReplanningThe process of dynamically updating or reconstructing an existing plan in response to execution failures or environment changes.
- ReplicateA cloud platform that allows developers to run, fine-tune, and deploy open-source machine learning models via an API.
- ReproducibilityThe ability to obtain consistent results using the same data, code, model parameters, and experimental setup.
- Request BatchingAn optimization technique that combines multiple incoming requests into a single batch to improve throughput and resource utilization.
- Request-ResponseA fundamental message exchange pattern where a client sends a request and a server returns a corresponding response.
- RerankingA retrieval process that re-scores and re-orders an initial list of candidate documents using a specialized model to improve relevance.
- Research AgentAn autonomous agent designed to gather, synthesize, and analyze information from various sources to answer complex research queries.
- Response CachingA performance optimization technique that stores previously generated responses to immediately serve identical future requests.
- Response SanitizationThe process of filtering or removing sensitive, unsafe, or non-compliant content from model outputs before they reach users.
- Response TimeThe total duration required for an AI system or service to process an input query and return the completed output.
- Responses APIAn API endpoint designed to manage, format, and return model-generated responses and computational outputs.
- Responsible AIA framework for developing and deploying AI systems ethically, safely, transparently, and in alignment with human values.
- RESTRESTAn architectural style for designing networked applications using stateless HTTP requests and standard operations.
- Retrieval OptimizationTechniques used to improve the accuracy, speed, and relevance of information retrieved from a knowledge base.
- Retrieval-Augmented GenerationRAGAn architectural pattern that enhances generative model responses by retrieving relevant information from external knowledge bases.
- RetrieverA system component responsible for fetching relevant documents or passages from a knowledge base given a query.
- Retry PolicyA policy that defines how and when a system re-attempts failed operations to handle transient errors.
- Retry StrategyA defined approach for automatically re-attempting failed operations or tasks to handle transient errors.
- Reward FunctionA mathematical function that quantifies the feedback or score an agent receives for taking a specific action in a given state.
- Reward ModelingThe process of training a mathematical model to score AI outputs based on human or automated preferences.
- Risk AssessmentThe process of identifying, evaluating, and mitigating potential safety, ethical, and operational risks in AI systems.
- Risk ThresholdA predefined limit or level of acceptable risk used to trigger safety interventions or operational escalations in AI systems.
- Risk-aware DecisionA choice made by evaluating potential outcomes alongside their associated uncertainties, hazards, and probabilities.
- RivetAn open-source visual programming environment and TypeScript library for building, testing, and debugging AI agent workflows.
- RobustnessA metric measuring an AI system's ability to maintain performance and stability under varied, unseen, or adversarial conditions.
- ROCmROCmAn open-source software platform developed by AMD for GPU computing and deep learning workloads.
- Role PromptingA prompting technique where an AI model is instructed to assume a specific persona or role to guide its responses.
- Role-Based Access ControlRBACA security model that restricts system access based on predefined roles assigned to individual users or agents.
- RollbackThe process of reverting an execution state or system to a previous known good state following an error or failure.
- Rollout StrategyA structured plan for deploying software or AI model updates safely to production while minimizing disruption.
- ROUGEROUGEA set of evaluation metrics used to measure the overlap of n-grams between generated text and reference summaries.
- Router ArchitectureAn agentic system design where a central router directs tasks and context to specialized agents based on request requirements.
- Rubric-based EvaluationAn evaluation approach that assesses AI outputs against structured, criteria-specific scoring rubrics.
- Rule EngineA software component that executes predefined logical rules to validate, filter, or control AI inputs and outputs.
- Runtime ContextThe active dynamic state, metadata, and environment parameters available to an agent or system during execution.
- Runtime EnvironmentThe underlying hardware and software infrastructure where applications or models execute.
- Runtime SpecificationA standardized blueprint or technical rule set defining the execution environment and lifecycle requirements for containers or workloads.
- Runtime StateThe dynamic status, variable values, and memory context preserved by an agent or system during active execution.
S
- Safe ActionAn operational state or behavior executed by an agent that satisfies defined safety constraints and safety policies.
- Safety FilterA mechanism designed to detect and block toxic, harmful, or policy-violating content in model inputs and outputs.
- Safety TestingThe process of evaluating AI systems to identify vulnerabilities, harmful outputs, and policy violations before deployment.
- Sandbox EnvironmentAn isolated runtime environment used to safely execute, test, and validate software or AI agent code without impacting production systems.
- ScaffoldingA technique for providing automated boilerplate structure, templates, or setup code to kickstart application development.
- Scenario-based EvaluationAn evaluation method that tests AI performance and behavior across realistic, multi-step scenarios or simulated environments.
- SchedulingThe process of allocating resources and assigning timestamps or execution order to planned tasks.
- Schema RegistryA centralized repository for storing, managing, and validating data schemas used across distributed systems.
- Schema ValidationThe process of verifying that structured data inputs or model outputs conform strictly to a predefined schema definition.
- SDK IntegrationThe process of embedding software development kits into an application to enable communication with external services or platforms.
- Secret ManagementThe processes and tools used to securely store, access, rotate, and manage sensitive credentials in software systems.
- Secret ManagerA secure tool or service designed to store, manage, and retrieve sensitive credentials and API keys in software applications.
- Security AssessmentA systematic evaluation of an AI system to identify security vulnerabilities, threats, and policy compliance gaps.
- Self-ConsistencySCA reasoning approach where multiple reasoning paths are sampled and the most consistent answer is selected as the final output.
- Self-ImprovementThe capability of an AI system to iteratively enhance its performance or reasoning using its own outputs, feedback, or data.
- Semantic CachingA caching technique that stores and retrieves query responses based on semantic similarity rather than exact string matches.
- Semantic KernelAn open-source SDK that allows developers to integrate AI services into traditional programming languages using plugins and prompts.
- Semantic MemorySMA long-term memory system that stores generalized facts, concepts, and world knowledge independently of specific experiences.
- Semantic SearchA search technique that analyzes sentence meaning and intent rather than matching literal keywords.
- Sensitive Data DetectionThe automated identification and scanning of confidential information or personally identifiable data within system inputs and outputs.
- Sequential PlanningA planning approach where tasks and actions are ordered and executed in a strict, step-by-step sequence.
- Sequential WorkflowA structured execution pattern where tasks are executed one after another in a linear sequence.
- Server-Sent EventsSSEA server-push technology enabling a client to receive automatic, real-time updates from a server over a single HTTP connection.
- Serverless DeploymentA deployment model where application code runs in ephemeral containers managed automatically by a cloud provider.
- Service ContractA formal specification defining the interface, input expectations, output structure, and protocols for a service.
- Service DiscoveryThe mechanism by which autonomous agents or network services automatically locate and identify available endpoints.
- Service EndpointA network address or communication entry point where a specific service or agent capability can be accessed.
- Service Level IndicatorSLIA quantifiable measure of the performance or health of a service used to determine compliance with service level objectives.
- Service ManifestA configuration document that defines the deployment properties, resources, and runtime environment of a service.
- Service MeshA dedicated infrastructure layer that manages service-to-service communication, security, and traffic routing across distributed applications.
- Service MetadataDescriptive information about a service's identity, capabilities, versions, and configuration attributes.
- Service OrchestrationThe automated arrangement, coordination, and management of complex software services and microservices to fulfill a workflow.
- Service RegistryA centralized database or dynamic repository that stores address and metadata information for available services and agents.
- Session ContextState and metadata maintained across interactions during an active communication session.
- Shadow TestingAn evaluation technique where live production traffic is duplicated to a candidate model or system without impacting the end-user response.
- Shared ContextContext information maintained across multiple agents or components during interaction.
- Shared ControlA collaborative paradigm where a human and an AI agent jointly manage and execute tasks or decision-making processes.
- Short-term MemorySTMA cognitive architecture layer that temporarily retains recent information and context for ongoing reasoning and execution.
- SidecarA structural design pattern where a secondary component attaches to a primary application to extend features without altering core code.
- Sidecar PatternA deployment pattern where a secondary container runs alongside a primary application to handle supporting features like logging or networking.
- Single AgentAn agent architecture where a standalone autonomous system operates independently without interacting with other agents.
- Single Sign-OnSSOAn authentication process that allows a user to access multiple applications or systems with a single set of login credentials.
- SmolAgentsA lightweight Hugging Face library designed for building AI agents using simple Python functions and minimal abstractions.
- Software Development KitSDKA collection of software development tools, libraries, and documentation that enables developers to build applications for a specific platform or service.
- Sparse RetrievalA search technique that uses term frequency and keyword matching, such as BM25, to locate relevant documents within a corpus.
- Starter TemplateA pre-configured project structure that provides baseline files and settings to accelerate application setup.
- Stateful WorkflowA process execution pattern that persists progress, context, and variables across steps to support long-running operations and recovery.
- Statistical SignificanceA mathematical determination that an observed experimental result is likely caused by a specific factor rather than random chance.
- Streaming ProtocolA network protocol designed for continuous real-time data transmission between clients and servers with minimal latency.
- Stress TestingAn evaluation technique that tests a system under extreme workloads or adversarial conditions to identify failure points.
- Structured PromptingA technique for designing prompts with explicit formatting rules or schemas to constrain and format model outputs.
- Success RateThe proportion of attempted tasks or requests that an AI agent or system completes successfully.
- Supervisor ArchitectureAn agent system design where a central supervisor agent coordinates, manages, and delegates tasks to subordinate specialized agents.
- Swarm IntelligenceA collective intelligence behavior emerging from decentralized, self-organized systems where simple agents interact locally.
- SWE-benchA benchmark for evaluating language models on resolving real-world software engineering issues from GitHub repositories.
- SynchronizationThe process of aligning states, operations, or data streams across multiple agents to maintain system consistency.
- System CardA detailed document outlining an AI system's architecture, intended use cases, safety evaluations, capabilities, and operational limitations.
- System PromptAn initial instruction that sets the baseline behavior, constraints, and operational context for a language model.
T
- Task DecompositionThe process of breaking down a complex task into smaller, manageable subtasks for structured execution.
- Task DelegationThe process where an AI agent assigns specific subtasks or responsibilities to other agents or systems.
- Task ExecutionThe process where an AI agent or system carries out a defined action, operation, or scheduled instruction.
- Task HandoffThe process of transferring operational control, execution context, and responsibility from one AI agent to another.
- Task PlanningThe process where an AI agent formulates an ordered sequence of actions or subtasks to achieve a given goal.
- Task RunnerA software tool that automates the execution of routine development tasks, build processes, and command sequences.
- Task-specific EvaluationAn assessment method that measures system performance on targeted, custom domain tasks rather than general capabilities.
- TCPTCPA core transport protocol that establishes reliable, ordered, and error-checked delivery of stream data between applications.
- TensorFlowAn open-source machine learning framework for building, training, and deploying deep learning models.
- TerraformAn open-source infrastructure as code tool used to define, provision, and manage cloud resources via declarative configuration files.
- Threat ModelingA systematic process for identifying, analyzing, and prioritizing potential security threats and vulnerabilities in a system.
- ThroughputThe amount of data, tokens, or requests processed by a system within a specified period.
- Time to First TokenTTFTThe duration of time it takes for a language model to produce its initial output token after receiving a request.
- TLSTLSA cryptographic protocol that provides end-to-end communication security and data encryption over computer networks.
- Together AIA cloud platform providing fast API access and infrastructure for hosting and running open-source AI models.
- Together AI SDKA software development kit provided by Together AI to interact with its cloud API for inference and model hosting.
- Token ExchangeAn OAuth 2.0 extension protocol for swapping security tokens to transfer identity and authorization context between services.
- Token MonitoringThe tracking and analysis of token usage and consumption patterns to manage costs and performance.
- Token OptimizationTechniques and strategies aimed at reducing prompt and completion token counts to lower latency and operational costs.
- Token UsageThe total volume of prompt and completion tokens consumed during interactions with a language model.
- TokenizersAn open-source library providing fast, production-ready tokenization for modern natural language processing models.
- Tokens per SecondTPSA performance metric measuring the rate at which an inference engine generates output tokens per second.
- Tool AdapterA software component that translates interface calls between external tools and an AI system for standardized execution.
- Tool CapabilityA specific action, function, or operational feature exposed by an external tool for an AI agent or system to execute.
- Tool ChainingThe practice of executing multiple tools sequentially, passing the output of one tool as input to the next.
- Tool ConnectorA middleware or software component that enables communication and data transfer between an AI system and an external tool.
- Tool DescriptorA standardized metadata record describing a tool's capabilities, input parameters, and execution schema.
- Tool DiscoveryThe mechanism by which an AI agent dynamically finds and inspects available tools and their capabilities at runtime.
- Tool EndpointA network address or API interface through which an AI system invokes a specific external tool or capability.
- Tool InvocationThe act of executing an external tool or function call with specified parameters as directed by an AI agent.
- Tool ManifestA document or file that specifies an external tool's metadata, available functions, and invocation requirements.
- Tool RegistryA centralized directory for storing, organizing, and discovering tools and their invocation specifications.
- Tool SelectionThe process by which an AI agent dynamically identifies and chooses the appropriate tool to perform a task.
- Tool TestingThe process of verifying that external tools, APIs, and function integrations behave correctly when invoked by an AI system.
- Tool ValidationThe process of verifying that external tool calls, arguments, and execution responses conform to expected schemas.
- Toxicity TestingThe process of evaluating an AI model or agent to detect and measure toxic, hateful, or abusive language in outputs.
- TPUTPUAn application-specific integrated circuit designed by Google to accelerate machine learning workloads.
- TransformersAn open-source Python library by Hugging Face providing pre-trained state-of-the-art transformer models.
- TranslatorA pattern that converts data formats, protocols, or schemas between incompatible systems to enable seamless communication.
- TransparencyThe degree to which an AI system's inner workings, decision logic, training data, and operations are visible and understandable.
- Tree of ThoughtsToTA reasoning framework where an AI system explores multiple branching thought paths to solve complex problems.
- Trial RunA preliminary execution of an experiment or pipeline designed to test functionality and validate configurations before full deployment.
- TRLTRLAn open-source Python library by Hugging Face for training transformer models using reinforcement learning techniques like RLHF and DPO.
- TruLensAn open-source software library for evaluating and tracking large language model applications and RAG pipelines.
- TruthfulQAA benchmark measuring how truthfully language models answer questions and avoid mimicking human falsehoods.
- Twelve-Factor AppA methodology for building software-as-a-service applications optimized for cloud deployment and scalability.
U
- UDPUDPA lightweight, connectionless transport protocol that transmits data packets across networks without establishing a prior connection.
- Unit TestingA software testing method that verifies the correctness of individual isolated units or components of code.
- Uptime MonitoringThe process of tracking and verifying the availability, functional operational status, and responsiveness of a system.
- User FeedbackDirect input provided by end users regarding their experience, satisfaction, or issues with an AI system's output.
- User PromptAn input message or instruction provided by a human user to direct an AI model's generation or task execution.
- User StudyAn evaluation method that assesses system performance and usability through direct observation and structured human interaction.
- Utility FunctionA mathematical function that assigns a numerical value to outcomes to represent preferences and guide optimal decision-making.
V
- Version ControlA system that tracks and manages changes to software code and assets over time.
- Vertex AI Vector SearchA fully managed Google Cloud vector database for high-scale, low-latency similarity search and retrieval.
- VespaAn open-source engine for storage, vector search, and real-time processing of large-scale structured data.
- Virtual Private CloudVPCAn isolated virtual network within a public cloud environment for hosting and securing cloud resources.
- vLLMAn open-source library for high-throughput and low-latency LLM serving and inference optimization.
W
- WeaviateAn open-source vector database designed to store data objects and vector embeddings for fast semantic search and AI applications.
- Web AgentAn AI agent designed to interact with web environments, perform tasks, and retrieve information across web applications.
- WebhookAn HTTP callback mechanism that sends real-time data payloads to target applications when specific events occur.
- Webhook IntegrationAn integration pattern that uses webhooks to deliver real-time data payloads and trigger automated actions between systems.
- WebRTCWebRTCA technology that enables real-time peer-to-peer audio, video, and data streaming directly between web browsers and devices.
- WebSocketA network protocol providing full-duplex, real-time communication channels over a single TCP connection.
- Weights & BiasesW&BAn MLOps platform for experiment tracking, model monitoring, dataset versioning, and AI system evaluation.
- Win RateThe proportion of evaluation tasks or head-to-head comparisons in which an AI model outperforms a baseline or competitor.
- Workflow OrchestratorA system that manages, coordinates, and executes complex multi-step workflows across distributed services.
- Workflow TestingThe process of verifying that multi-step AI tasks and execution flows operate correctly, reliably, and as intended.
- Working MemoryWMA cognitive system in AI architectures that temporarily holds and manipulates information required for active reasoning and task processing.
- WrapperA software component that encapsulates an interface or service to provide a simplified or unified API access layer.
X
- xAIAn artificial intelligence company that develops frontier AI models and platforms including Grok.
- XMLXMLA markup language designed to store and transport structured data in a human-readable and machine-readable format.
- XML PromptingA technique that uses XML tags to structure prompts and separate instructions, context, and data for language models.
- XML-RPCXML-RPCA remote procedure call protocol that uses XML to encode its calls and HTTP as a transport mechanism.
Y
Z
Signal, not noise.
Focused newsletter for builders and knowledge workers tracking how AI is changing real work. We surface what matters in practice, not every headline. Curated for practitioners, not spectators.