# Glossary

This glossary defines the key terms used throughout our documentation: models, API concepts, deployment options, and platform features. Each entry includes a short definition and links to the relevant documentation page. Use the search bar or letter filters to find a specific term.

- **Agent** (Vibe, API): An AI assistant configured with specific instructions, tools, and an optional knowledge base. Agents behave consistently across conversations and can be shared across a team. In Vibe, the equivalent feature in Work is called a Skill. The legacy Chat tab is gone for most users, so use Skills in Work instead. Agents can also be built programmatically via the Agents API in Studio.
- **Context window** (Concepts): The maximum number of tokens a model can process in a single request, including input (prompt and files) plus output combined. Earlier content is dropped when the limit is exceeded. → Token
- **Embeddings** (Models, Concepts): Numerical vector representations of text that encode semantic meaning. Similar text produces similar vectors, enabling similarity search, clustering, and classification. Mistral Embed produces 1024-dimensional vectors. Used in: RAG, search, recommendations.
- **FIM** (Models, Concepts): (fill-in-the-middle) A code generation mode where the model generates text based on both preceding (prefix) and following (suffix) context. Used for mid-function completion in IDEs. Available in Codestral.
- **Free mode** (Billing): Default Mistral access mode. No credit card required. Includes monthly usage with rate limits suitable for evaluation and prototyping. → Pay-as-you-go
- **Function calling** (API, Concepts): (tool use) A capability that lets the model request execution of external functions or APIs mid-conversation. Your application executes the function and returns the result, allowing the model to incorporate real-world data into its response.
- **Guardrailing** (Concepts): Policies applied at the system level to restrict or shape model output, enforcing content policies, defining permitted output formats, and setting behavioral constraints without modifying the underlying model.
- **Inference** (Concepts): Running a trained model to generate a response. Calling the API is running inference, as distinct from training, which modifies model weights.
- **OCR** (API, Concepts): (optical character recognition) Extracting structured text and images from document files. Mistral OCR processes PDFs and scanned documents, preserving reading order and layout structure.
- **Pay-as-you-go** (Billing): Optional usage billing beyond included monthly usage. Billed by token with no monthly commitment. Suitable for workloads above Free mode limits. → Free mode
- **RAG** (Concepts): (retrieval-augmented generation) A technique that improves factual accuracy by retrieving relevant documents from an external knowledge base and including them in the prompt context. Reduces hallucinations and enables models to answer from proprietary or up-to-date data. Used with: Embeddings, Agents.
- **Streaming** (API, Concepts): A response delivery mode where tokens are sent as they are generated rather than as a single complete response. Reduces perceived latency in interactive applications such as chat interfaces.
- **Studio** (Platform): Mistral AI's SaaS API and developer platform at console.mistral.ai. Provides API access, the Studio developer console, and model management tools.
- **System prompt** (Concepts): An instruction at the start of a conversation that shapes model behavior throughout the session. Defines persona, output format, constraints, and context. Not shown to end users, but consumes tokens toward the context window limit.
- **Temperature** (API, Concepts): A parameter (0.0 - 1.0) that controls output randomness. Lower values (0.2) produce focused, deterministic output. Higher values (0.8) produce more varied and creative output. Set to 0 for tasks where consistency and reproducibility are critical.
- **Token** (Billing, Concepts): The basic unit of text models process: subword fragments, not full words. Roughly 1 token ≈ 0.75 English words ≈ 4 characters. Token count determines API pricing and context window consumption.
- **Vibe** (Vibe): Mistral's unified agent. Runs in two modes: Work (the unified chat and work experience in the web and mobile apps at chat.mistral.ai) and Code (CLI, VS Code extension, or remote web sessions at code.mistral.ai). Legacy Chat features such as Agents and Memories have equivalents in Work, and the legacy Chat tab remains available only to some Enterprise organizations during the migration window.
- **Workspace** (Platform, Billing): An isolated environment within a Mistral organization. Each workspace has its own API keys, usage metrics, and billing bucket. Use workspaces to separate development and production, or to isolate usage and spending by team or project.
