AI Glossary

An AI agent combines a language model (GPT-4, Claude, Gemini) with a set of tools (function calling) to execute tasks. Unlike a chatbot, which answers questions, an agent acts: queries databases, sends emails, modifies a CRM, calls APIs, runs code and plans multiple steps to solve a task. Example: a support agent that receives an incident, looks up information in your knowledge base (RAG), checks the customer's history in the CRM, proposes a solution, executes it if authorised and logs the interaction.

A Large Language Model is an AI model trained on hundreds of billions of words (books, web, code) that learns language patterns and world knowledge. The most used in production are GPT-4 / GPT-4o (OpenAI), Claude Sonnet / Opus (Anthropic), Gemini (Google) and open-source models such as Llama (Meta) or Mistral. Capabilities: generate text, translate, summarise, code, reason and — connected to tools — execute complex tasks as agents.

RAG works in two steps: (1) when a question arrives, the system retrieves the most relevant fragments of your documents using vector search over embeddings; (2) those fragments are passed as context to the LLM, which writes the answer based on them. Benefits: accurate and citable answers, always-up-to-date data (just re-index) and a drastic reduction of hallucinations. Typical use cases: support chatbots over corporate documentation, sales assistants with up-to-date catalogue and pricing, semantic search across knowledge bases.

A traditional chatbot follows predefined flows (decision trees) and answers with templates. Modern chatbots use LLMs and understand natural language without a script. Combined with RAG, they answer based on your documentation. Add function calling and tools and they become AI agents that can execute actions. Common channels: web, WhatsApp Business, Telegram, Messenger, Microsoft Teams, Slack.

An AI automation goes beyond a traditional workflow (Zapier, Make, n8n): it includes an LLM or agent to make contextual decisions, write responses, classify tickets, extract data from unstructured documents, etc. Typical examples: automatic classification of incoming emails, lead response generation, data extraction from scanned invoices, daily report generation from sales data, automated support during off-hours.

An embedding is a vector of hundreds or thousands of dimensions generated by a model (OpenAI text-embedding-3, Cohere, etc.) that represents the semantic meaning of a text. Two texts similar in meaning have close embeddings in vector space, even if they don't share words. Example: "book a table" and "make a reservation at the restaurant" have very similar embeddings. It's the foundation of RAG, semantic search, support ticket clustering and duplicate detection.

A vector database stores vectors (embeddings) and finds the closest ones to a given query in milliseconds. The most used: Pinecone, Weaviate, Qdrant, Milvus, Chroma and the pgvector extension for PostgreSQL. For small cases (up to ~100k vectors), pgvector on top of your existing Postgres is more than enough. For larger scale (millions), Pinecone or Qdrant are the typical choice.

Prompt engineering is about how to phrase instructions (system prompt, user prompt, few-shot examples) so an LLM produces useful, consistent and safe answers. Best practices: be specific about the output format, give examples, set explicit constraints, use chain-of-thought ("think step by step") in reasoning tasks. In serious projects prompts are versioned and tested against an evaluation dataset.

Fine-tuning takes a pretrained model (GPT-4o-mini, Llama, Mistral) and retrains it on your own dataset (typically between 50 and thousands of question-answer pairs) to specialise it. Useful when you need a very specific response style, guaranteed structured JSON output or niche knowledge. Most projects do NOT need fine-tuning: with good prompt engineering and RAG it's usually enough. Cost: typically €50–€300 for a fine-tune that ships to production.

MCP defines a JSON-RPC standard for models to access tools (functions), resources (data) and predefined prompts. Instead of building custom integrations every time, you expose an MCP server and any compatible client (Claude Desktop, Cursor IDE, your own agents) can use it. Examples: MCP server for your CRM, for a file system, for a database, for Google Workspace. We implement it when your data/systems need to be available both for external assistants and for in-house AI agents.

Function calling lets the model decide which function in your system to execute (query_customers, send_email, search_product) and with what arguments, based on the conversation. The runtime executes the function, returns the result to the model and the model writes the final answer. Supported by GPT-4, Claude, Gemini and open-source models. It's the foundation on which AI agents and MCP are built.

A multimodal model understands and generates more than one type of content. It can analyse a PDF invoice + photo, transcribe audio, describe a video or generate text from a diagram. Typical enterprise uses: extracting data from scanned documents, automatic product photo analysis, meeting transcription and summarisation, customer service via WhatsApp voice notes. Models: GPT-4o (text+image+audio), Claude Sonnet (text+image), Gemini Pro (text+image+video+audio).

Hallucinations happen when the model generates plausible but invented data (names, figures, references to laws). Mitigations: (1) RAG with mandatory source citations, (2) prompts that require "don't make things up, say I don't know", (3) downstream validation with tools (function calling to query real data), (4) human in the loop for critical decisions. In production projects we treat hallucinations as bugs and build automatic evaluations to detect them.

A token is typically a short word or part of a word. "Hello, how are you?" is ~6 tokens. Models charge per token (input and output) and have a maximum context window: GPT-4o has 128k, Claude Sonnet 200k, Gemini 1.5 Pro up to 2M. For long documents you need to split into chunks or use long-context models. Typical cost: 1M input tokens costs between $0.15 (mini) and $15 (top model).

Instead of a single generalist agent, a multi-agent system splits the tasks: a "receptionist agent" understands the request and routes it, a "researcher agent" looks up information, a "writer agent" prepares the answer and a "reviewer agent" validates. Useful for complex flows where a single prompt would be too long or ambiguous. Frameworks: LangGraph, CrewAI, AutoGen, custom systems with bespoke orchestration.

Closed models are usually more capable but you send your data to a third party (with zero-data-retention clauses in enterprise use). Open-source models you self-host on your own infrastructure: maximum privacy, no per-token cost, but it requires GPU and maintenance. Typical decision: start with a closed API (fast, cheap to test) and migrate to open source when volume or privacy justify it. Llama 3.3 70B and Qwen 2.5 72B come close to GPT-4 on many tasks.

Reasoning models generate explicit chains of thought before the final answer, significantly improving on maths, logic, programming and complex planning. They are slower and more expensive than fast models (GPT-4o, Claude Haiku) but indispensable for tasks that require deep analysis: process optimisation, hard debugging, legal or strategic analysis. Typical use: combine a reasoner for planning with a fast model for execution.