Anvia

@anvia/mistral API reference ​

Import every public symbol from @anvia/mistral. The package has no public subpath exports.

Client and model factories ​

ts
type MistralClientOptions =
  | { apiKey: string; baseUrl?: string; client?: never }
  | { client: Mistral; apiKey?: never; baseUrl?: never }

class MistralClient implements ModelListingClient {
  constructor(options: MistralClientOptions)
  completionModel(options: MistralCompletionModelOptions): MistralCompletionModelHandle
  embeddingModel(options: MistralEmbeddingModelOptions): MistralEmbeddingModelHandle
  ocrModel(options: MistralOcrModelOptions): MistralOcrModelHandle
  listModels(options?: { abortSignal?: AbortSignal }): Promise<ModelList>
}

type MistralCompletionModelOptions = {
  modelId: MistralCompletionModelId
  contextLimits?: ModelContextLimits
}

type MistralEmbeddingModelOptions = {
  modelId: MistralEmbeddingModelId
  dimensions?: number
  maxBatchSize?: number
}

type MistralOcrModelOptions = { modelId: MistralOcrModelId }

All constructors and factories require one options object. Managed credentials and an injected Mistral SDK client are mutually exclusive. Model IDs are explicit; there are no factory defaults.

The completion and embedding factories return the Core StreamingCompletionModel and EmbeddingModel contracts. OCR returns a handle exposing provider, modelId, and ocr().

OCR ​

MistralOcrSource accepts document_url, image_url, file_id, or bytes. Byte sources require a filename and may set provider expiry and visibility. MistralOcrRequest also accepts page selection, image extraction limits, annotations, table format, header/footer extraction, confidence granularity, provider options, and an abort signal.

MistralOcrResponse includes combined text and markdown, normalized pages, optional model and usage data, optional upload metadata, and the raw provider response. Empty byte sources are rejected.

Mapping helpers ​

ts
function toMistralChatParams(
  modelId: MistralCompletionModelId,
  request: CompletionRequest,
): Record<string, unknown>

function fromMistralChatResponse(response: unknown): CompletionResponse
function fromMistralChatStreamChunk(chunk: unknown): CompletionModelStreamEvent[]

mistralMessageHelpers separately exposes messageToMistralMessages() and toolDefinitionToMistral(). These low-level exports use the same normalization as the completion handle.

Model IDs and namespace ​

Completion, embedding, and OCR IDs use exported ModelId<Known...ModelId> types. Known IDs provide autocomplete while other strings remain valid. MISTRAL_OCR_LATEST is also exported.

The mistral namespace re-exports the same client, helpers, constants, and public types. It has no state of its own.

Built for Anvia.