Understand artificial intelligence for the Web
AI vocabulary often mixes model capability, access to documents and ability to act through tools. Separate these dimensions. A fluent answer can be incorrect; a connected tool can perform an action without the model understanding every consequence.
32 concepts to explore with definitions and practical advice.
Search the full glossary
A reading path
1. Define the task
Specify question, allowed sources and expected result. A prompt cannot replace missing information or an undefined business rule.
- LLM
Large Language Model trained on large volumes of text to understand and generate language.
- Prompt
Instruction, context or set of examples given to an AI model to guide its response or action.
- Grounding
Mechanism connecting a model answer to retrieved data, documents or results to improve accuracy and verifiability.
2. Understand context and tools
Document retrieval adds context; a tool protocol organises exchanges with applications. Define separately what the system can read and what it can change.
- Retrieval-Augmented Generation
Technique retrieving relevant external information before asking a model to generate an answer grounded in that context.
- Model Context Protocol
Open protocol allowing an AI application to discover and use tools or data sources exposed in a structured way.
- AI agent
AI system able to chain steps, use tools and pursue a goal with a degree of autonomy.
3. Evaluate difficult cases
Keep authorised examples and success criteria. Test missing sources, conflicting information and out-of-scope requests before deploying the journey.
- AI evaluation
Reproducible protocol measuring the quality, safety, consistency or cost of an AI system on defined cases.
- Hallucination
Confidently produced answer containing invented, distorted or insufficiently supported information.
- Algorithmic transparency
Information provided about an algorithmic system’s purpose, data, limits, decisions and governance.
Useful distinctions
Retrieval-Augmented Generation / Model Context Protocol
RAG names generation grounded in retrieved information. MCP organises access to compatible resources and tools. Neither automatically requires the other, and neither guarantees accuracy.
Chatbot / AI agent
A chatbot is a conversational interface. An agent can receive goals and use tools within its scope. The agent label does not remove the need to define permissions, limits and action controls.
Use the vocabulary to make a decision

Situation
An assistant gives a convincing answer whose links do not support its facts.
Before deciding
Check claims individually against authorised documents. A displayed citation does not establish that the source actually says what is claimed.
A concrete check
Include this case in the evaluation set and plan an explicit response when information cannot be confirmed.
Evaluate an answer against verifiable evidence
Prepare examples matching the real task: an available answer, a missing source and conflicting documents. Decide beforehand how the system should handle each case.
- ClaimWhich fact does the answer assert?
- DocumentWhich passage actually supports it?
- DecisionApprove, correct or request clarification?
On a narrow screen, scroll the table horizontally to read all three columns.
| Observed signal | What to check | What it means |
|---|---|---|
| A citation opens but does not support the claim | Compare each claim with the exact passage. | The presence of a link does not establish the result. |
| Two documents give different conditions | Compare dates, versions and scopes. | The system should report the conflict rather than invent agreement. |
| The useful source is missing | Test an answer that acknowledges this limitation. | An explicit incomplete answer may be preferable to invented certainty. |
Next step : Separate factual, scope and action errors; repeat the cases after each system change.
Technical reference : NIST — AI Risk Management Framework. The scenarios and numerical examples above are educational, not measurements of this website.
A cited reference still needs checking
Fictional teaching case. Open each interpretation to examine its scope; there is no score or professional assessment.
- Claim isolated
- Passage retrieved
- Scope checked
An assistant produces an answer with a source URL. The page title seems relevant, but nobody has checked whether its content supports the stated claim.
What evidence is available?
A useful next step : Separate generated text, a retrieved source and a verified fact. Keep disagreements and unconfirmed information visible before publication.
Reference to explore : NIST · AI Risk Management Framework · Consulted .
All concepts in this topic
Links open the full definition on its alphabet page. Acronyms and synonyms remain searchable in the main glossary.
- Agentic commerce
- Agentic Web
- AI Act
- AI agent
- AI citation
- AI crawler
- AI evaluation
- AI-assisted search
- Algorithmic transparency
- Artificial intelligence
- Chatbot
- Chunking
- Embedding
- Generative AI
- Grounding
- Hallucination
- LLM
- Machine learning
- Model Context Protocol
- Model routing
- Multimodal model
- Natural language processing
- Neural network
- OAI-SearchBot
- Prompt
- Prompt engineering
- Retrieval-Augmented Generation
- Token
- Transformer
- Vector database
- Vector search
- Web agent
Frequently asked questions
Does Web access eliminate errors?
No. It can provide new information, but selection, reading and interpretation still require checking. Keep the primary source and context.
Should tools be connected for the first trial?
Start with capabilities required by the task. Add a tool when usefulness and limits are defined, then evaluate parameter errors and out-of-scope requests.
Primary reference documents
These documents explain technical concepts. Reading paths and examples are editorial methods to adapt to your project.