Glossary of chatbots & conversational AI agents.

A large-scale language model, trained on vast amounts of text, that generates answers by predicting the next word. It is the engine that produces a conversational assistant's language. Examples used in implementations: GPT-4o, Claude, Llama, Mistral.

Appeared

The Transformer architecture — 2017 (Google); consumer LLMs — from 2022 (ChatGPT).

In implementation

We pick the model per case: GPT-4o for conversational reasoning, Claude for long documents and nuanced tone, open-source models (Llama, Mistral) for on-premise or tight budgets.

A technique where, before answering, the assistant searches the relevant fragments from your documents and hands them to the model as context. The answer is thus anchored in real data, not in the model's generic memory — the primary method for reducing hallucinations.

Appeared

Concept formalized in 2020 (Facebook AI / Meta paper), a standard in business assistants since 2023.

In implementation

We split documents into fragments, turn them into embeddings, store them in a vector database and, on every question, inject the matching fragments into the prompt.

The collection of sources the assistant "knows" — product catalog, prices, FAQs, manuals, internal policies. It is the source of truth from which every answer is built. The cleaner and more complete it is, the more precisely the assistant answers.

Appeared

A pre-AI customer-support concept; redefined as a source for RAG from 2023.

In implementation

We import catalog, documents, FAQs and manuals; we structure them and refresh them periodically. The Starter package covers up to 50 documents; Business and Enterprise — unlimited knowledge base.

The representation of a text as a list of numbers (a vector) that captures meaning, not the exact words. Texts with close meaning have close vectors. It enables semantic search: the assistant finds the relevant fragment even when the user uses different words than the document.

Appeared

Word2Vec — 2013; sentence/document-level embeddings went mainstream from 2022.

In implementation

We turn every fragment of the knowledge base into embeddings and compare them with the question's embedding to retrieve the most relevant passages.

A database optimized for storing and quickly searching embeddings. It answers the question "which fragments are semantically similar to this query?" in milliseconds, even across millions of documents. The "long-term memory" component of a RAG system.

Appeared

A category that consolidated in 2022–2023 (Pinecone, Weaviate, pgvector, Qdrant).

In implementation

We choose the solution by volume and hosting requirements; for EU/on-premise deployments we use self-hosted options, GDPR-compliant.

Retraining a model on your own examples so it adopts a specific style, format or behavior. Unlike RAG (which adds knowledge at answer time), fine-tuning changes the model itself. Useful for tone and repetitive tasks; more expensive to maintain.

Appeared

A standard transfer-learning practice; applied commercially to LLMs from 2023.

In implementation

Reserved for the Enterprise package, when the brand needs a dedicated model. For most cases, RAG + a well-written system prompt achieve similar results, more cheaply.

The set of instructions, invisible to the user, that fix the assistant's role, tone, rules and limits before any conversation. This is where the persona is defined, what it may and may not say, and how it escalates. The backbone of behavior.

Appeared

A standard mechanism of chat LLMs from 2022.

In implementation

Here we encode the persona, guardrails and the fallback rules. A poorly written system prompt yields a generic assistant; a well-written one often removes the need for fine-tuning.

The maximum amount of text (measured in tokens) a model can process in a single exchange — including the instructions, the conversation history and the injected documents. When the conversation exceeds it, the model "forgets" the oldest part unless a memory mechanism exists.

Appeared

A concept from the earliest LLMs; windows grew sharply from 2023.

In implementation

We manage the window by summarizing history and through selective RAG, so only the relevant context fits — keeping latency low.

The assistant's distinct personality — tone, vocabulary, signature lines, the way it reacts to jokes and objections. It turns an animated FAQ into an interlocutor that speaks like the brand. The difference between a generic chatbot and an assistant that actually sells lies, first of all, in the persona.

Appeared

A conversation-design discipline that solidified alongside generative AI assistants, 2023–2024.

In implementation

We define the persona in a workshop with the team (Day 1): voice, target audience, objective. Live examples: Dona (ironic, pharma), Hanna (consultative, HVAC), Sierra (adventurous, premium auto).

A conversation that unfolds across several successive turns, where the assistant understands references to earlier messages and fills missing information by asking naturally. The opposite of a bot that treats each message in isolation and forgets what was said one turn ago.

Appeared

A standard capability of LLM assistants from 2022.

In implementation

The assistant keeps the thread: if you ask "and for menopause?" after a discussion about supplements, it understands what you mean without you restating everything.

The mechanism by which the assistant retains relevant information throughout a conversation (short-term memory) or across sessions (long-term memory). It enables continuity: the user need not repeat context, and answers stay coherent from one turn to the next.

Appeared

Memory techniques for conversational agents, mature from 2023.

In implementation

We combine the recent conversation history with RAG over profile/history, keeping only what is relevant so we don't fill the context window.

The real goal behind a message — what the user actually wants, beyond the words used. "How much is it?", "do you have discounts?" and "is it expensive?" can share the same intent. Reading intent correctly determines whether the assistant answers the question asked or the one missed.

Appeared

A central concept in classic chatbots (from ~2016); reinterpreted by LLMs, which infer it without rigid rules.

In implementation

In modern assistants, intent is inferred by the model in context, not hardcoded into a decision tree — hence the flexibility on nuanced questions.

A system's ability to understand natural language — the meaning, intent and entities in a freely written message, with typos, slang or unexpected phrasing. The component that separates an assistant that "gets it" from a button menu that demands exact wording.

Appeared

An NLP subfield with a long history; revolutionized by LLMs from 2022.

In implementation

In AI assistants, NLU comes natively from the LLM: we no longer train separate intent classifiers, relying instead on the model's contextual understanding.

An answer that sounds plausible and self-assured but is false or invented — a wrong price, a non-existent policy, a fictional source. The biggest business risk of an AI assistant: a hallucination destroys trust exactly when the customer was ready to buy.

Appeared

A term popularized alongside consumer LLMs, 2022–2023.

In implementation

We counter it through RAG (anchoring in real data), grounding, clear knowledge limits and fallback rules: when it doesn't know, the assistant says so and escalates. It never invents.

The practice of forcing the assistant to build its answers exclusively from verifiable sources — your documents — and, ideally, to cite where the information comes from. A "grounded" answer can be verified; an ungrounded one is just a well-phrased guess from the model.

Appeared

A principle that solidified with the production adoption of RAG, 2023.

In implementation

We configure the assistant to answer from the knowledge base, not from generic knowledge, and to explicitly acknowledge when a piece of information is missing.

The rules that set what the assistant may and may not do — forbidden topics, mandatory tone, the duty not to invent, not to pose as human, to escalate in certain situations. The "red line" that protects the brand from an assistant gone off the rails.

Appeared

An AI-safety discipline applied commercially from 2023.

In implementation

Our red lines: we don't build assistants without personality, we don't allow lying/hallucination, the assistant always states it is AI, and we process data only in a GDPR-compliant way.

What the assistant does when it reaches the limit of its knowledge or competence. Instead of inventing, it runs a predefined plan: it says it doesn't know and suggests human contact, automatically escalates to a live agent, or collects contact details and promises a follow-up.

Appeared

A conversation-design concept, essential in the age of hallucinations.

In implementation

We define the exact fallback variant per brand, in the Discovery phase. Good fallback is invisible to the customer — it reads as care, not failure.

The transfer of the conversation from the AI assistant to a human agent, keeping all context intact — the history, the question and the collected data. The customer doesn't start over. It triggers when the conversation exceeds the AI's scope or when sentiment indicates frustration.

Appeared

A standard feature in customer-engagement platforms; integrated with AI from 2023.

In implementation

Available from the Business package. The transfer is instant and complete — the agent picks up exactly where the assistant left off.

Collecting contact details through natural conversation (not a dry form) and qualifying the prospect by your criteria — budget, need, timing — before passing them to the sales team. The assistant turns curious visitors into qualified quote requests, with no human effort.

Appeared

A marketing-automation practice reinvented conversationally from 2023.

In implementation

The assistant asks naturally, qualifies by your rules and sends the lead straight to the CRM. Available from the Business package.

Connecting the assistant to the customer-relationship management system, so leads, conversations and sentiment flow automatically to where the team works. It eliminates manual copy-paste and ensures no prospect generated by the assistant is lost.

Appeared

Native AI–CRM integrations, mature from 2023–2024.

In implementation

We connect HubSpot, Pipedrive, Salesforce, Monday or a custom API (Enterprise). The lead, the transcript and the score land in the right pipeline.

Automatic detection of the conversation's emotional tone — satisfied, neutral, frustrated. It lets the assistant adapt its answer and escalate to a human before an unhappy customer leaves. In aggregate, it shows which topics generate recurring frustration.

Appeared

A classic NLP technique, built into modern conversational dashboards.

In implementation

Sentiment feeds the handoff decision and lands in the analytics dashboard as actionable data — not dead reports.

The time elapsed between the user's question and the assistant's first response. High latency kills the conversation: the customer perceives the hesitation as a malfunction. The practical goal for a business assistant is a response perceived as almost instant.

Appeared

A universal performance metric; critical for conversational UX.

In implementation

We optimize latency per channel; the operational target is a response under 3 seconds. The choice of model and management of the context window are the main levers.

The place where the assistant lives and talks to customers — a website widget, WhatsApp Business, Facebook Messenger, Instagram Direct. The same persona, the same knowledge, on any channel. For the Romanian market, WhatsApp is often the preferred channel for communicating with brands.

Appeared

An omnichannel strategy adapted to AI assistants, 2023–2024.

In implementation

An embeddable widget on any site technology (no re-engineering), plus WhatsApp Business API, Messenger and Instagram from the Business package. 24/7 response, the same voice everywhere.

The European legal framework governing how the personal data of a conversational assistant's users is collected, processed, stored and deleted. It covers the place of processing, encryption, explicit consent and the retention period. For serious companies, compliance is not optional.

Appeared

GDPR in force since 25 May 2018; applied to conversational AI systems from their adoption.

In implementation

We process data in the EU, with end-to-end encryption and explicit consent at the start of the session. Retention is configurable. We don't use your data to train public models.