mistralai

Enables synchronous and asynchronous text generation across Mistral's model lineup (Mistral 7B, Mistral 8x7B, Mistral Large, Mistral Small) via a unified client interface that abstracts model selection and handles both complete responses and token-by-token streaming through iterator patterns. The SDK manages request serialization, response deserialization, and connection pooling to the Mistral API endpoints.

Implements tool/function calling by accepting JSON schema definitions of available functions, sending them to Mistral models with user prompts, and parsing structured responses that indicate which function to call with what arguments. The SDK handles schema validation, response parsing, and provides helper methods to map function names back to callable Python functions for execution.

Provides a Message class hierarchy (UserMessage, AssistantMessage, ToolMessage) that structures multi-turn conversations with role-based semantics, enabling the SDK to maintain conversation context across API calls. The client accepts a list of messages and automatically formats them for the API, handling role validation and message ordering without requiring manual serialization.

Implements both synchronous and asynchronous client classes (MistralClient and AsyncMistralClient) using httpx for HTTP transport, enabling concurrent API calls via Python's asyncio event loop. The async client supports streaming responses through async generators, allowing non-blocking token consumption in event-driven applications.

Provides an embeddings API endpoint that converts text input into fixed-dimensional dense vectors using Mistral's embedding models. The SDK handles text chunking, batch processing, and returns embedding vectors as lists of floats, enabling semantic search and similarity computations without external embedding services.

Automatically extracts and returns metadata from API responses including token counts (prompt tokens, completion tokens, total tokens), model identification, and finish reasons (stop, length, tool_calls). This metadata is attached to response objects, enabling cost tracking and quota management without additional API calls.

Defines custom exception classes (MistralAPIError, MistralConnectionError, etc.) that wrap HTTP errors and API-specific failures, providing structured error information including status codes, error messages, and retry hints. The client automatically raises these exceptions on API failures, enabling granular error handling without parsing raw HTTP responses.

Exposes a list_models() method that queries the Mistral API to discover available models, their capabilities, and metadata (context window, max tokens, etc.). This enables dynamic model selection and capability checking without hardcoding model names, supporting applications that adapt to available models.

Allows configuration of HTTP timeout values, connection pool sizes, and retry behavior at client initialization. The SDK passes these settings to httpx, enabling fine-grained control over network behavior without modifying SDK code. Timeout configuration applies to both streaming and non-streaming requests.

Handles API key authentication by accepting a Mistral API key at client initialization and automatically injecting it into request headers. The SDK supports environment variable loading (MISTRAL_API_KEY) and explicit key passing, enabling flexible credential management without hardcoding secrets in code.