Key Concepts of Generative AI

Updated November 2, 2026 — Le Chat becomes Vibe. As of May 28, 2026, Le Chat is now Vibe (still at chat.mistral.ai): a single agent for work and code, carrying over your conversations, settings, and subscriptions. The legacy Le Chat experience remains accessible in Chat mode; in Work mode, agents are now skills (invoked with @), memories became the knowledge base, and deep search became the Deep Research skill. Plans are now called Mistral Free, Pro, Team, and Enterprise. This lesson now uses the name Vibe; some menus described here may have changed name or location.

Mastering the essential vocabulary

There are a handful of terms you’ll come across in every meeting, every piece of documentation, and every discussion with a technical team about generative AI: RAG, embeddings, fine-tuning, temperature, hallucinations, grounding. This lesson explains them one by one, not as glossary definitions, but as tools you’ll know how to place in a real conversation. When a vendor proposes “a RAG chatbot with Mistral embeddings,” you’ll know exactly what they’re selling you—and what questions to ask.

RAG — Retrieval-Augmented Generation

RAG (retrieval-augmented generation) is one of the most widespread techniques for improving the quality of an LLM’s responses. The principle starts from a simple observation: instead of relying solely on the model’s internal knowledge, you provide it, at the time of the question, with relevant information extracted from an external database.

The mechanism unfolds in four steps:

1. The user asks a question
2. A search system identifies relevant documents in a knowledge base
3. These documents are injected into the model’s context
4. The model generates its response based on this fresh information

Why is this so important? Because RAG solves a fundamental problem with LLMs: their knowledge is frozen at the date of their training. A model trained last year knows nothing about your new product catalog, the reorganization of your department, or the regulations published last month. With RAG, it can answer questions about recent events, about data internal to your company, or about documents you provide. Imagine an HR department receiving the same questions every week about the collective bargaining agreement: a RAG assistant connected to company agreements answers by citing the correct article, instead of improvising from memory. This is the technique that powers Vibe’s web search and most enterprise chatbots.

Embeddings — Vector representations

Embeddings are numerical representations of text in the form of vectors, meaning lists of numbers. Each word, sentence, or document is transformed into a point in a high-dimensional mathematical space. The key idea fits in one sentence: two semantically close texts will have embeddings that are close in this space. “Dog” and “canine” will be neighbors; “dog” and “spreadsheet” will be very far apart, even though neither of these pairs shares significant common letters.

This property changes everything for information retrieval. A traditional keyword search won’t find a document about “mutual termination” if you search for “amicable contract ending”; a semantic search by embeddings will, because both formulations point to the same area of the vector space. Beyond search, embeddings are used to automatically classify texts by theme, detect duplicates or similar content, and—as we saw in the previous lesson in this section—to select relevant documents in a RAG pipeline. Mistral offers dedicated embedding models, Mistral Embed and Codestral Embed, which transform your text into usable vectors.

Fine-tuning — Tailored adaptation

Fine-tuning consists of continuing the training of a pre-trained model with your own data, to adapt it to a specific need. The most accurate image is that of a general practitioner trained as a specialist: the general knowledge is already there, you add pointed expertise.

You resort to it in several situations. A brand that wants all its customer responses to respect a specific tone and style can fine-tune a model on its best past exchanges. A law firm or medical department can instill industry-specific vocabulary that the generalist model handles roughly. A team processing thousands of emails per day can specialize a model on a specific task like classification or entity extraction, with better results than a generic prompt. Finally, distillation consists of transferring knowledge from a large model to a small one, to achieve similar performance at a much lower cost. At Mistral, it’s the Forge platform that lets you fine-tune Open models: you prepare a dataset of question-answer pairs, you launch the training, and you get a customized model deployable via the API.

Temperature — The creativity slider

Temperature is a parameter that controls the degree of “creativity”—technically, randomness—in the model’s responses. At low temperature (0 to 0.3), the model systematically chooses the most probable responses: the result is predictable, precise, and reproducible, which suits factual tasks, data extraction, and code. At medium temperature (0.4 to 0.7), you get a balance between coherence and variety that suits most conversational uses. At high temperature (0.8 to 1.0), the model explores less obvious formulations: more creativity, but also more risk of errors—a setting to reserve for brainstorming and creative writing.

On Vibe, temperature is managed automatically; you don’t have to think about it. Via the API however, you can define it precisely for each request, and it’s a powerful lever: the same question asked at 0.1 and at 0.9 will produce an accountant’s response in one case, a copywriter’s response in the other.

Hallucinations — When the model makes things up

Hallucinations are responses that seem plausible but are factually incorrect. The model “invents” information, and—here’s the danger—it does so with the same confident tone as when it tells the truth. The cause is structural: an LLM predicts probable tokens, not verified facts. It can therefore produce text that is impeccable linguistically and false factually. Risk areas are well identified: precise dates, figures and statistics, citations attributed to real people, references to studies or articles, and specific technical details. Ask a model for the bibliography of a little-known author: it will willingly provide you with book titles that are perfectly credible… that do not exist.

You don’t eliminate hallucinations, you limit them. RAG anchors responses in verified sources; a low temperature favors the most probable responses; asking the model for sources—and then verifying them yourself—reveals inventions; grounding, which we see just after, connects the model to real data. And on any critical subject, the rule remains absolute: never blindly trust a response.

Grounding — Anchoring in reality

Grounding refers to all the techniques that connect a model to reliable and up-to-date information sources. RAG is one form, but the concept is broader. Web search lets the model go verify facts on the Internet rather than draw from its frozen memory. Connection to databases gives it access to your internal data—your actual inventory, your actual customers. Function calling, finally, lets it call external APIs to get real-time data: the weather, a stock price, the status of a system. The difference is decisive: grounding moves from an assistant that “knows things” to an assistant that “can verify things.”

Prompt — The instruction to the model

The prompt is the text you send to the model: your instruction, your question, your request. Its quality largely determines that of the response—it’s the subject of an entire lesson later in this course. Keep three notions in mind now. The system prompt groups the general instructions given to the model upfront (tone, role, constraints), configurable via the API. The user prompt is your actual message, the question or task at hand. And few-shot prompting consists of providing examples of expected responses to guide the model: showing is often better than explaining.

Key takeaways

• RAG enriches the model’s responses with fresh external data
• Embeddings transform text into vectors for semantic search
• Fine-tuning adapts a model to your specific needs via Forge
• Temperature controls the balance between precision and creativity
• Hallucinations are plausible but false responses; grounding anchors the model in verifiable data to limit them