Glossary

A practical, plain-English glossary of AI, chatbot, and platform terms used across EchoWiseAI. These definitions are original and tailored to how our product works so your team shares the same vocabulary.

A

Agent

An AI-driven workflow that can decide what to do next, call tools or APIs, and iterate toward a goal. In EchoWiseAI, agents combine models, tools, and knowledge.

Artificial Intelligence (AI)

Techniques that let software perform tasks associated with human intelligence: understanding language, reasoning, and decision-making.

AI Model

A statistical program (often a large language model) that generates or classifies text from input and prior training.

Action (Custom Action)

A callable operation the assistant can execute (e.g., call your REST API, query a database). Exposed to the model via tool/function calling.

Analytics

Usage and performance metrics for conversations, costs, and deflection used to improve the assistant.

API

A programmatic interface that lets systems communicate. We expose APIs for embedding and management; we also call your APIs via Actions.

API Key

A secret token used to authenticate to an API. Store securely and rotate when needed.

B

Bot (Chatbot)

A conversational interface that responds to user messages via text or voice.

Bot Persona

The tone, style, and behavior guidelines that shape how your assistant communicates.

C

Cache

Temporary storage used to speed up repeated operations (e.g., retrieving embeddings or knowledge lookups).

Chunking

Splitting documents into smaller passages before indexing so retrieval returns focused, relevant snippets.

Conversation

A sequence of messages between a user and the assistant, grouped by a session identifier.

Context

Information supplied to the model (system messages, examples, retrieved passages) that guides how it responds.

Context Window

The maximum number of tokens a model considers at once, including the prompt and the generated output.

Crawling

Automatically following links on a site to discover pages to ingest. Different from scraping, which extracts a specific page.

Custom Domain

A domain you configure so the widget and assets appear under your brand (e.g., chat.yourdomain.com).

D

Data Retention

How long conversation data and documents are stored. Configurable based on compliance or business needs.

DPA (Data Processing Addendum)

A contract defining how personal data is processed on your behalf to meet privacy laws.

Determinism

The degree to which a model returns the same output for the same input. Lower temperature increases determinism.

Document

A unit of knowledge you add (HTML page, PDF, CSV, FAQ). Documents are chunked, embedded, and indexed for retrieval.

E

Embedding

A numerical vector representation of text used for semantic search and similarity comparisons.

Entity

A specific piece of information (e.g., order number, email, product name) extracted from a message.

Entity Extraction

Identifying entities in user messages to power workflows, analytics, or validation.

Evaluation

Measuring answer quality, safety, and latency using test prompts and real conversations.

Export

Downloading conversation transcripts, leads, or knowledge data for audits or migration.

F

Fallback

A safe, generic reply when the assistant cannot confidently answer; can route to a human or suggest alternatives.

Fine‑tuning

Training a base model further on your examples to specialize behavior. Often paired with RAG for factual accuracy.

Function/Tool Calling

Letting the model request that the runtime execute a specific function or API with structured arguments, returning results into the conversation.

G

GDPR

EU privacy regulation governing how personal data is collected, stored, and used.

Grounding

Supplying authoritative context (retrieved passages, policies) so answers stay factual and on‑brand.

Guardrails

Rules and filters that constrain model behavior, enforce safety, and maintain tone.

H

Hallucination

A confident but incorrect response from a model. Reduced by retrieval, citations, and good prompts.

Human‑in‑the‑Loop (HITL)

A workflow where humans review or take over conversations, provide labels, or approve sensitive actions.

I

Index

The searchable structure built from embeddings and metadata that powers fast retrieval.

Ingestion

The process of importing, cleaning, chunking, and embedding your content for search.

Intent

The purpose behind a user’s message (e.g., track order, cancel subscription, pricing question).

Integration

A prebuilt connection to third‑party systems (e.g., Shopify, WooCommerce, Slack) used by actions or data syncs.

J

Job Queue

Background workers that handle long‑running tasks like crawling, training, and analytics aggregation.

K

Knowledge Base

The collection of documents and data sources that your assistant can reference during retrieval.

L

Latency

The time between a user message and the assistant’s first token of response. Streaming reduces perceived latency.

Lead

Contact information captured in chat (e.g., email, phone) for follow‑up. Configurable with consent prompts.

Large Language Model (LLM)

A model trained on large text corpora to predict and generate natural language.

M

Max Tokens

The limit on how many tokens the model may generate in a single response.

Memory (Conversation Memory)

Information about the current user and prior messages that the assistant can recall within a session.

Metadata

Descriptive attributes stored with each chunk or message (e.g., source URL, timestamp) to control retrieval and analytics.

Moderation

Automated checks for unsafe or non‑compliant content before or after generation.

Multimodal

Models or chats that handle more than text (e.g., images, audio, or structured outputs).

N

Natural Language Processing (NLP)

Methods that let computers understand, interpret, and generate human language.

Nucleus Sampling (Top‑p)

A decoding strategy that samples from the smallest set of tokens whose cumulative probability is at least p.

O

Omnichannel

A unified experience across multiple channels (web widget, email, messaging apps) with consistent context and identity.

Orchestration

Coordinating prompts, retrieval, tools, and post‑processing to deliver a final answer.

P

Persona

The set of instructions that define the assistant’s voice, audience, and do/don’t rules.

PII (Personally Identifiable Information)

Data that can identify a person (e.g., name, email, IP). Handle according to privacy laws and your policies.

Prompt

The full input sent to a model: system instructions, user message, examples, and retrieved context.

Prompt Engineering

The craft of shaping prompts and inputs to get consistent, useful outputs from models.

Provider (Model Provider)

The company offering a model API (e.g., OpenAI, Anthropic). You can switch providers per use case.

Q

Quality Assurance (QA)

Systematic checks—often with test sets and human review—to ensure responses meet standards.

Queue

A system for processing tasks asynchronously to improve reliability and throughput.

R

Retrieval‑Augmented Generation (RAG)

A pattern where the model first retrieves relevant knowledge chunks, then uses them to craft an answer—improving accuracy and reducing hallucinations.

Rate Limit

The maximum number of requests allowed over a time window. Exceeding limits returns 429 errors.

Redaction

Masking or removing sensitive data (e.g., emails, credit cards) from logs and outputs.

Retrieval

Finding the most relevant chunks from your knowledge base using embeddings, keywords, or hybrid search.

S

Safety Filters

Policies and automated checks that block disallowed content and reduce harmful outputs.

Schema

A structured definition of expected inputs/outputs (often JSON) used for tool calling or structured generation.

Scraping

Programmatically extracting content from a web page. Use with permissions and respect robots rules.

Semantic Search

Searching by meaning using embeddings, rather than exact keyword matches.

Session

A bounded interaction period tied to a user or device. Sessions group messages and memory.

Shopify Integration

A connection that lets the assistant look up orders, products, and policies in your Shopify store.

Similarity Score

A numeric measure of how closely two embeddings match; higher means more similar.

Sitemap

An XML file that lists site URLs for discovery and ingestion.

Streaming

Sending tokens as they are generated so users see the answer appear in real time.

Stripe

The billing platform used for subscriptions, usage metering, and invoices.

Structured Output

Asking the model to return data in a strict format (e.g., JSON) for reliable downstream handling.

T

Temperature

A randomness control. Lower values yield focused, repeatable responses; higher values are more creative.

Token

A piece of text (roughly ~4 characters in English) used by models. Prompts and outputs are measured in tokens.

Tokenization

The process of splitting text into tokens that a model can process.

Tool

A function the model can call to fetch data or perform actions (e.g., check order status, create a ticket).

Top‑k

Sampling from the k most likely next tokens during generation.

Top‑p

See Nucleus Sampling.

Turn

One exchange of user message and assistant reply.

Throttling

Slowing or limiting requests to protect services and respect quotas.

U

Utterance

Any single user input sent to the assistant.

Usage

The count of tokens, messages, or requests consumed by a workspace, model, or user over time.

V

Vector

A list of numbers representing the meaning of text in an embedding space.

Vector Database

A storage and search system optimized for embeddings and nearest‑neighbor queries.

Visitor

An anonymous user interacting with your widget before identifying themselves.

W

Webhook

An HTTP callback we send to your system when events occur (e.g., lead captured, conversation finished).

Widget

The chat interface you embed on your site or app via a small script tag.

X

CORS (Cross‑Origin Resource Sharing)

Browser rules controlling which domains can make requests to your APIs or load your assets.

Z

Zero‑shot / Few‑shot

Techniques where the model performs a task with no examples (zero‑shot) or with a small number of examples (few‑shot) in the prompt.

Want a term added? Contact support and we’ll expand this glossary to match your use cases.