Distinctionary

Un Distinctionary combina un glossario consultabile con una visualizzazione di tassonomia gerarchica, permettendoti di esplorare la terminologia degli agenti IA sia attraverso la ricerca diretta che le relazioni concettuali.

Termine	Definizione
A/B testing	Comparing variants under controlled conditions to choose the best performing approach.
Access control	Mechanisms like RBAC that restrict who can run models, view data, or invoke tools.
Action	An operation taken by the agent, such as calling a tool, retrieving data, or generating content.
Adapter	A small parameter module inserted into a model to enable efficient task adaptation.
Agent architecture	A category for concepts that define how agents plan, act, observe, and manage state over time.
Agent types	A category grouping classical and modern agent forms by complexity and capabilities to guide architecture selection.
Agentic AI	An approach where autonomous or semi‑autonomous agents carry out end‑to‑end tasks using planning, tooling, memory, and verification loops.
Agentic workflow	A coordinated process where one or more agents plan, act, and verify to complete a task or case.
AGI	Artificial general intelligence with human‑level performance across most cognitive tasks, a debated future goal.
AI agent	A software entity that senses inputs, selects actions, and optimizes behavior to achieve objectives in an environment.
AI Agents	An umbrella concept for software systems that perceive, decide, and act toward goals with varying autonomy, encompassing simple agents through multi‑agent, tool‑using, and memory‑equipped systems.
AI risk management	Coordinated activities to direct and control risks associated with AI systems and their use.
Alignment	Ensuring model goals and behaviors are consistent with human values and policies.
Annotation	Human or tool labeling that creates training or evaluation data for models.
Anonymization	Irreversibly removing personal identifiers so individuals cannot be reidentified.
API integration	Connecting agents to external services via authenticated, schema‑defined endpoints.
ASI	Artificial superintelligence that would surpass human capabilities across domains, a speculative concept.
Assistant message	Model or agent responses that continue the conversation and may include tool calls.
Attention	A mechanism that lets the model focus on the most relevant parts of the input when producing outputs.
Audio embedding	A vector representation of audio useful for search and classification.
Audit trail	Tamper‑evident logs that record actions for compliance and forensics.
Backoff	Increasing wait times between retries to reduce contention and protect services.
Barge-in	A feature that lets a caller interrupt prompt playback with speech or DTMF to speed interactions.
Barge-in handling	Policies and logic to safely process interruptions without losing dialogue state.
Benchmark	A standard task suite for comparing systems under consistent conditions.
Bias	Systematic error introduced by data, modeling choices, or deployment context that skews outputs.
BM25	A classic probabilistic ranking function for lexical retrieval commonly used as a baseline.
Caching	Reusing prior results to reduce latency, cost, and tool pressure for repeated queries.
Call orchestration	Coordinating ASR, NLU, business tools, and TTS across the call lifecycle for quality outcomes.
Canary prompt	A sensitive prompt used to detect regressions in safety or reasoning after changes.
Canary release	Deploying a change to a small slice of traffic to validate safety and quality before full rollout.
Chain-of-thought prompting	Eliciting intermediate reasoning steps to improve solution quality on complex tasks.
Chunking	Splitting documents into appropriately sized passages to improve retrieval precision and context fit.
Citation	A reference to a source that supports an output statement for verification.
Compliance	Meeting legal, regulatory, and policy requirements in AI development and operations.
Confidence score	A numeric or categorical estimate of output reliability for ranking or escalation decisions.
Consent	Explicit user permission for data collection or processing in specific contexts.
Content filtering	Blocking or transforming unsafe or off‑policy content before it reaches users.
Context awareness	Using history and situational data to tailor responses to the user's current environment and needs.
Context length	The token capacity of the model's input window that bounds how much can be considered.
Context stuffing	Overfilling the prompt with too much or irrelevant context that can degrade output quality.
Context window	The maximum token span a model can consider at once, functioning like short‑term memory.
Conversation history	The prior messages that inform current context and continuity.
Cost budget	The maximum allowed spend in tokens or currency for a task or session.
Cost per task	Total spend in tokens and tools required to complete a task successfully.
Cost tracking	Measuring token, tool, and infrastructure spend at task and system levels.
Data governance	Policies and controls that ensure responsible data use across the AI lifecycle.
Data minimization	Collecting and retaining only data necessary for a specified purpose.
Data retention	Policies that define how long data is kept and when it is deleted.
Deduplication	Removing duplicate or near‑duplicate items to reduce noise and bias in retrieval or training data.
Delegation	Assigning a subtask to another specialized agent or a human to improve efficiency and outcomes.
Differential privacy	Adding noise to data or queries to protect individual privacy while enabling aggregate analysis.
Document store	A repository for unstructured or semi‑structured documents used by retrieval systems.
DPIA	Data Protection Impact Assessment to evaluate and mitigate risks to individuals' rights from processing.
Drift	Degradation due to changes in data, user behavior, or external tools over time.
DTMF	Dual‑Tone Multi‑Frequency signals from keypad input used for IVR navigation and data entry.
Dynamic resource allocation	Adjusting model size, tools, or retrieval depth at runtime to meet task complexity and constraints.
Embeddings	Numerical vector representations of text or other data that preserve semantic similarity for search and clustering.
Emergent behavior	Capabilities that arise in large models without explicit training for those skills.
Encryption	Encoding data at rest and in transit so only authorized parties can read it.
Evaluation	Systematic measurement of agent quality, robustness, cost, and safety using tasks and metrics.
Evaluation and metrics	A category for measuring agent quality, robustness, cost, and safety.
Evaluation harness	Reusable tasks, datasets, and metrics for consistent measurement across iterations.
Event	A notable change such as a tool result or user input that can trigger agent logic.
Evidence	Specific retrieved passages or data points used to substantiate an output or decision.
F1 score	The harmonic mean of precision and recall providing a single effectiveness measure.
Failure modes taxonomy	A categorization of common agent failure patterns to inform controls and testing.
Feedback loop	Collecting and incorporating outcomes to improve future model or agent performance.
Few-shot prompting	Supplying a few labeled examples in the prompt to guide the model to the desired task behavior.
Fine-tuning	Updating a pretrained model's weights on domain data to specialize behavior.
Foundations	A category for core model, prompting, and training concepts that underpin agent capabilities.
Function calling	Structured tool invocation where the model outputs a function name and arguments for execution.
Functional agent	A deterministic integration‑oriented agent that bridges AI prompts to APIs or devices with minimal autonomy or learning.
GDPR	EU regulation governing personal data processing, rights, and cross‑border transfers.
Glossary	A controlled vocabulary that standardizes definitions across teams and systems.
Goal-based agent	An agent that selects actions by evaluating how choices advance explicitly defined goals.
Governance frameworks	A category for structured practices to manage AI risk, accountability, and security across agents.
Gradient	The derivative of the loss with respect to parameters used to update the model during training.
Ground truth	Authoritative labels or facts used to train or evaluate models and agents.
Grounding	Linking outputs to cited sources or retrieved evidence to ensure factuality and traceability.
Guardrails	Controls that prevent unsafe, off‑policy, or low‑quality outputs during agent execution.
Handoff latency	The time required to transfer a session from an agent to a human while preserving context.
Hierarchical agent	An agent organized in tiers where higher levels decompose tasks and supervise lower‑level execution.
HIPAA	A U.S. healthcare regulation protecting medical information and governing its use and disclosure.
Human-in-the-loop	Humans review or intervene in critical steps to ensure safety, accuracy, and accountability.
Hybrid search	Combining lexical (e.g., BM25) and semantic retrieval to improve recall and precision.
Idempotency	Design that ensures repeated requests have the same effect as one, preventing duplicates.
In-context learning	Adapting behavior from examples and instructions in the prompt without weight updates.
Incident response	Coordinated steps to triage, mitigate, and learn from production failures or breaches.
Index	A data structure that accelerates search over documents or vectors for fast retrieval.
Instructions	Rules and objectives given to a model or agent that shape behavior and output quality.
IVR	Interactive Voice Response systems that interact with callers via menus, speech, or DTMF.
Jailbreak	A technique to circumvent safety controls and elicit restricted outputs from a model.
Key management	Secure generation, storage, rotation, and use of cryptographic keys.
Knowledge and retrieval	A category for grounding methods that connect agents to external knowledge.
Knowledge base	A curated repository of documents and answers used to support retrieval and support agents.
Knowledge graph	A graph of entities and relationships that supports reasoning and semantic retrieval for agents.
Large language model	A transformer‑based model trained on large corpora to understand and generate language and code.
Latency	The time from request to response, often tracked at p50, p95, and p99 for reliability.
Latency target	A specified real‑time responsiveness goal for voice interactions (e.g., <300 ms turn‑start).
Latent space	The continuous vector space where models represent features and relationships of data.
Learning agent	An agent that adapts behavior from feedback or data to improve performance over time.
Logging	Recording inputs, actions, tool calls, and outputs for debugging, audits, and evaluation.
Long-term memory	Persistent knowledge stored outside the context window and retrieved when relevant to the task.
LoRA	A lightweight fine‑tuning technique that learns low‑rank updates to model weights.
Loss function	A scalar objective that measures prediction error during training and guides learning.
Max tokens	The maximum number of tokens the model can generate in the response.
Memory	Stored information from prior interactions or facts that an agent can recall for continuity and accuracy.
Model	A parameterized system that maps inputs to outputs based on learned patterns from data.
Model versioning	Tracking and controlling model variants and their metadata across environments.
Model-based reflex agent	A reflex agent that maintains an internal model of the world to infer hidden state while still lacking planning.
Monitoring	Continuous measurement of agent behavior, latency, cost, and quality during operation.
Multi-agent system	A system of multiple interacting agents that coordinate or compete to achieve complex objectives.
Multimodal	Models or agents that process and generate across text, images, audio, and other modalities.
NIST AI RMF	A framework with core functions Govern, Map, Measure, Manage to manage AI risks and improve trustworthiness.
Observability	Using logs, metrics, and traces to understand internal system behavior from external outputs.
Observation	Feedback received after an action, such as tool results or user responses, used to update state.
Operations and SRE	A category for running agents in production with reliability, cost, and change management controls.
Orchestration	Coordinating models, prompts, tools, data, and control flow to execute AI tasks reliably and repeatably.
Orchestration layer	The component that sequences prompts, tools, memory, and control logic for agents.
Overfitting	When a model memorizes training data patterns and fails to generalize to new data.
OWASP agentic taxonomy	Security‑oriented taxonomy and patterns for building and testing secure autonomous agents.
Persona	A defined role and tone that constrain how an agent communicates and prioritizes actions.
PII	Personally identifiable information that requires heightened protection and access control.
Plan	A structured, ordered set of steps that guides the agent through execution toward a goal.
Planning	Creating a sequence of steps or subgoals that an agent follows to accomplish a task.
Playbook	Guided patterns and templates for implementing repeatable AI workflows.
Policy	A set of rules and preferences that constrain agent behavior for safety, compliance, and quality.
Policy compliance	Adhering to organizational, legal, and platform rules in prompts, tools, and outputs.
Positional encoding	Signals added to tokens so a transformer can represent order in sequences.
Postmortem	A blameless analysis after incidents to identify root causes and corrective actions.
Precision	The fraction of returned items that are relevant, used to assess retrieval or classification.
Privacy	Protecting individuals' data and limiting use to authorized, disclosed purposes.
Privacy and compliance	A category for safeguarding data and meeting regulatory obligations across the lifecycle.
Profile	Structured attributes about a user or entity that personalize an agent's decisions and responses.
Prompt	Input text (and other signals) that instructs a model on task, style, and constraints.
Prompt injection	A malicious input that manipulates the model to ignore instructions or perform disallowed actions.
Prompting and sampling	A category for techniques that steer generation and control randomness and diversity.
Provenance	Metadata describing where data and outputs came from and how they were produced.
Pseudonymization	Replacing identifiers with tokens while retaining linkability under controlled conditions.
Query expansion	Improving retrieval by adding related terms or reformulations to the original query.
Queue	A buffer that evens out bursty workloads and supports retries and ordering.
Rate limit budget	The allotted quota of calls or tokens the agent can consume over time.
Rate limiting	Restricting request frequency to control load, cost, and fair use of APIs.
Re-ranking	Reordering retrieved results with a secondary model to improve relevance at the top of the list.
Real-time monitoring	Observing latency, errors, and quality during active sessions for rapid intervention.
Reasoning	The process of transforming inputs and knowledge into intermediate thoughts that justify and guide actions.
Recall	The fraction of relevant items that are returned, complementing precision.
Red teaming	Adversarial testing of AI systems to find failures, safety gaps, and abuse pathways.
Reflection	Having the agent review its own steps or outputs to correct errors and improve future actions.
Regularization	Techniques that reduce overfitting by discouraging overly complex or brittle models.
Retries	Re‑issuing failed calls with safeguards to improve reliability without runaway costs.
Retrieval	Selecting relevant items from a corpus based on a query to inform downstream generation.
Retrieval-Augmented Generation (RAG)	Generating answers grounded in external knowledge fetched at query time to reduce hallucinations.
Retry policy	A strategy for repeating failed calls with limits and delays to improve robustness.
Runbook	Step‑by‑step procedures for handling common operational tasks and issues.
Safety	Practices and controls that reduce harm, misuse, and unintended behaviors in AI systems.
Safety and governance	A category for practices and controls that reduce harm and ensure responsible use.
Safety policy	Documented rules describing allowed and disallowed content and behaviors.
Safety violation rate	The frequency of outputs that breach safety policies over a defined sample.
Scheduler	A component that runs tasks or checks triggers at specified times or intervals.
Secret management	Managing credentials and tokens required for tool and API access.
Self-attention	Computing attention within a sequence so each token can use information from other tokens.
Semantic search	Finding relevant content by meaning rather than exact keyword matches using embeddings.
Session	A bounded interaction period that groups turns, state, and budgets for an experience.
Short-term memory	The active context available inside a model's context window during the current turn or task.
Simple reflex agent	A rule‑based agent that reacts to current percepts without internal state or planning.
SLA	A contractual commitment on availability and performance targets for a service.
Sliding window	A strategy that feeds overlapping context segments to a model to fit long inputs within limits.
SLO	An internal target for key reliability metrics such as latency and uptime.
Smoke test	A lightweight check that validates basic functionality and safety before full runs.
SOC 2	A service organization control standard for security, availability, processing integrity, confidentiality, and privacy.
Source citation	Attaching references to outputs so users can verify origin and relevance of information.
Source grounding	Confining answers to retrieved or known sources to improve factuality and trust.
Speech-to-text	Automatic conversion of spoken language into written text.
State	The internal information an agent maintains across steps, such as goals, plans, and intermediate results.
System prompt	The highest‑priority instructions that define role, goals, and boundaries for the assistant or agent.
Task decomposition	Breaking a complex goal into simpler, solvable subtasks to improve reliability and control.
Task success rate	Share of tasks completed to specification without human escalation.
Telemetry	Collected metrics and traces that describe agent performance and errors in production.
Temperature	A randomness parameter controlling output diversity, with higher values yielding more varied text.
Termination criteria	Conditions that signal an agent to stop, such as achieving goals, hitting limits, or detecting failures.
Terms of use	Conditions governing how users and organizations may access and use an AI system.
Text-to-speech	Synthesis of natural‑sounding audio from text.
Throughput	The number of tasks or tokens processed per unit time under given resources.
Timeout	The maximum wait time before abandoning a tool call or model request.
Token	The smallest unit of text the model processes, such as words or subword pieces.
Token usage	The number of tokens consumed across prompts, contexts, and generations.
Tokenization	Converting text into tokens according to a model's vocabulary rules.
Tool catalog	The set of available tools with schemas and permissions that an agent can invoke.
Tool invocation	Executing an external function with structured inputs produced by the model.
Tool output	The result returned by an invoked function that the agent uses as new context.
Tool schema	A formal specification of function name, parameters, and types for safe calling.
Tool selection	Choosing which available tools to call and in what order based on the task and context.
Tool use	Selecting and invoking specific tools during a task to extend an agent's capabilities beyond text generation.
Tool-augmented retrieval	Using tools or APIs during retrieval to transform queries or fetch structured data beyond static corpora.
Tools	External functions or APIs an agent can call to obtain information or take real‑world actions.
Top-k	Sampling that restricts token choices to the top k most probable candidates.
Top-p	Nucleus sampling that limits choices to the smallest set whose cumulative probability exceeds p.
Training and optimization	A category for methods that adapt or improve models for tasks and domains.
Transcription	Converting speech audio into text with ASR models and pipelines.
Transformer	A neural architecture built on self‑attention that powers modern LLMs.
Trigger	A condition or signal that initiates a workflow step or agent action.
Truncation	Dropping least‑relevant context when inputs exceed the model's token limit.
Trustworthy AI glossary	A NIST resource that standardizes terminology to align teams on responsible AI concepts.
Turn count	The number of conversational exchanges needed to resolve a user's request.
User prompt	User‑provided instructions or questions that the model responds to within constraints.
Utility-based agent	An agent that optimizes a numeric utility function to balance trade‑offs among competing outcomes.
Variance	The sensitivity of a model's predictions to changes in training data, often trading off with bias.
Vector database	A database optimized for storing and searching embeddings via nearest‑neighbor similarity.
Voice activity detection	Detection of speech segments in audio to control turn‑taking and barge‑in.
Voice agent	An agent that communicates over telephony or voice channels with ASR and TTS in real time.
Voice and multimodal	A category for real‑time speech, telephony, and non‑text modalities in agent interactions.
Webhook	An HTTP callback that delivers event notifications to trigger workflows.
Windowing	Techniques that slide or segment long inputs to fit within the context window.
Workflow	An ordered set of steps and decision points that structure how tasks are completed.
Zero-shot prompting	Performing a task from instructions alone without any examples in the prompt.