ai-provider-mistral-sdk

Mistral SDK Patterns

Quick Guide: Use @mistralai/mistralai (ESM-only) to interact with Mistral's API. Use client.chat.complete() for chat, client.chat.stream() for streaming (async iterable via for await), client.chat.parse() with a Zod schema for structured outputs, and client.fim.complete() for Codestral fill-in-middle code completion. The SDK uses responseFormat (camelCase) not response_format. Streaming events expose content via event.data.choices[0]?.delta?.content. Retries default to strategy: "none" -- you must configure them explicitly for production.

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use responseFormat (camelCase) in SDK calls -- NOT response_format (snake_case). The SDK uses camelCase property names throughout.)

(You MUST configure retries explicitly -- the SDK defaults to strategy: "none" (no retries), unlike OpenAI's SDK which retries automatically)

(You MUST consume streaming results with for await (const event of result) and access content via event.data.choices[0]?.delta?.content -- the event shape differs from OpenAI)

(You MUST never hardcode API keys -- use process.env["MISTRAL_API_KEY"] with the bracket notation the SDK documents)

(You MUST use client.chat.parse() with a Zod schema for structured outputs -- NOT manual JSON.parse() on completion content)

</critical_requirements>

Auto-detection: Mistral, mistral, @mistralai/mistralai, client.chat.complete, client.chat.stream, client.chat.parse, client.fim.complete, client.embeddings.create, mistral-large, mistral-small, codestral, pixtral, ministral, magistral, devstral, MISTRAL_API_KEY, responseFormat, mistral-embed

When to use:

• Building applications that call Mistral models directly (Mistral Large, Small, Codestral, etc.)
• Implementing chat completions with SSE streaming
• Using Codestral for code generation and fill-in-middle (FIM) completion
• Extracting structured data with client.chat.parse() and Zod schemas
• Implementing function calling / tool use
• Creating embeddings for RAG pipelines or semantic search
• Processing images with vision-capable models (Mistral Small, Medium, Large, Ministral)
• Using Mistral Agents API for pre-configured agent completions

Key patterns covered:

• Client initialization and configuration (retries, timeouts, custom HTTP client)
• Chat completions (chat.complete) and streaming (chat.stream)
• Structured outputs with chat.parse() and Zod schemas
• Function calling / tool use with tool call loop
• Embeddings (embeddings.create) with mistral-embed
• Vision (image URL / base64 with vision-capable models)
• Codestral FIM (fim.complete) for code completion
• Error handling, retry configuration, and production patterns

When NOT to use:

• Multi-provider applications where you need to switch between Mistral, OpenAI, Anthropic, etc. -- use a unified provider SDK
• React-specific chat UI hooks (useChat) -- use a framework-integrated AI SDK
• When you need OpenAI-compatible endpoints -- use OpenAI SDK with Mistral's compatible endpoint instead

Examples Index

• Core: Setup & Configuration -- Client init, production config, error handling, retries, custom HTTP client
• Chat & Streaming -- Chat completions, streaming with async iteration, multi-turn
• Structured Output -- chat.parse() with Zod, JSON mode, typed responses
• Function Calling -- Tool definitions, tool call loop, streaming tools
• Embeddings & Vision -- Semantic search, image analysis with vision-capable models
• Codestral FIM -- Fill-in-middle code completion, code generation
• Quick API Reference -- Model IDs, method signatures, error types, configuration options

Philosophy

The @mistralai/mistralai SDK is auto-generated from Mistral's OpenAPI spec using Speakeasy, giving you a thin, type-safe wrapper over the REST API. It is ESM-only and uses camelCase property names (not snake_case like the REST API).

Core principles:

1. ESM-only -- The package is published as ESM only. CommonJS projects must use await import(). This is a hard constraint, not optional.
2. camelCase API surface -- SDK properties use camelCase (responseFormat, maxTokens, toolChoice) even though the REST API uses snake_case. This catches OpenAI SDK migrants who write response_format.
3. No automatic retries -- Unlike OpenAI's SDK (2 retries by default), Mistral defaults to strategy: "none". You must configure retries explicitly for production.
4. Streaming via async iterables -- chat.stream() returns an EventStream consumed with for await...of. Events have a data wrapper: event.data.choices[0]?.delta?.content.
5. Structured outputs via chat.parse() -- Pass a Zod schema directly to responseFormat and access message.parsed for typed results. No manual JSON schema construction needed.
6. Codestral FIM -- Dedicated fim.complete() endpoint for fill-in-middle code completion, separate from chat.

When to use the Mistral SDK directly:

• You only use Mistral models and want the simplest, most direct integration
• You need Mistral-specific features (Codestral FIM, Mistral Agents, Voxtral audio)
• You want minimal dependencies and zero abstraction overhead
• You need the latest Mistral API features on day one

When NOT to use:

• You need to switch between providers (OpenAI, Anthropic, Mistral) -- use a unified provider SDK
• You want React-specific chat UI hooks -- use a framework-integrated AI SDK
• You want an OpenAI-compatible wrapper -- Mistral exposes an OpenAI-compatible endpoint, use the OpenAI SDK for that

Core Patterns

Pattern 1: Client Setup

Initialize the Mistral client. It reads MISTRAL_API_KEY from the environment.

// lib/mistral.ts -- basic setup
import { Mistral } from "@mistralai/mistralai";

const client = new Mistral({
  apiKey: process.env["MISTRAL_API_KEY"] ?? "",
});

export { client };

// lib/mistral.ts -- production configuration
import { Mistral } from "@mistralai/mistralai";

const TIMEOUT_MS = 30_000;

const client = new Mistral({
  apiKey: process.env["MISTRAL_API_KEY"] ?? "",
  timeoutMs: TIMEOUT_MS,
  retryConfig: {
    strategy: "backoff",
    backoff: {
      initialInterval: 1_000,
      maxInterval: 30_000,
      exponent: 1.5,
      maxElapsedTime: 120_000,
    },
    retryConnectionErrors: true,
  },
});

export { client };

Why good: Explicit retry config (SDK defaults to no retries), named constants, env var with bracket notation

See: examples/core.md for custom HTTP client, async API key provider, error handling

Pattern 2: Chat Completions

Basic chat using chat.complete().

const result = await client.chat.complete({
  model: "mistral-large-latest",
  messages: [
    { role: "system", content: "You are a helpful coding assistant." },
    { role: "user", content: "Explain TypeScript generics." },
  ],
});

const content = result?.choices?.[0]?.message?.content;
console.log(content);

Why good: Uses system role for instructions, safe optional chaining on nullable response

// BAD: Using snake_case properties (REST API style, not SDK style)
const result = await client.chat.complete({
  model: "mistral-large-latest",
  messages: [{ role: "user", content: "hello" }],
  response_format: { type: "json_object" }, // WRONG: use responseFormat
  max_tokens: 100, // WRONG: use maxTokens
});

Why bad: SDK uses camelCase properties -- response_format and max_tokens will be silently ignored

See: examples/chat.md for multi-turn, token tracking, temperature control

Pattern 3: Streaming

Use chat.stream() for streaming. Events are async iterables.

const result = await client.chat.stream({
  model: "mistral-large-latest",
  messages: [
    { role: "system", content: "You are a helpful assistant." },
    { role: "user", content: "Explain async/await in TypeScript." },
  ],
});

for await (const event of result) {
  const content = event.data.choices[0]?.delta?.content;
  if (content) {
    process.stdout.write(content as string);
  }
}
console.log();

Why good: Proper for await iteration, accesses event.data (not event directly), handles nullable delta

// BAD: Trying to access content directly on event (OpenAI pattern)
for await (const chunk of result) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? ""); // WRONG
}

Why bad: Mistral streaming events wrap data in event.data -- direct access on chunk will fail

See: examples/chat.md for complete streaming examples

Pattern 4: Structured Outputs with Zod

Use chat.parse() with a Zod schema for type-safe structured responses.

import { Mistral } from "@mistralai/mistralai";
import { z } from "zod";

const client = new Mistral({ apiKey: process.env["MISTRAL_API_KEY"] ?? "" });

const BookSchema = z.object({
  name: z.string(),
  authors: z.array(z.string()),
});

const MAX_TOKENS = 256;

const result = await client.chat.parse({
  model: "mistral-large-latest",
  messages: [
    { role: "system", content: "Extract the book information." },
    { role: "user", content: "I recently read 'Dune' by Frank Herbert." },
  ],
  responseFormat: BookSchema,
  maxTokens: MAX_TOKENS,
  temperature: 0,
});

const parsed = result.choices?.[0]?.message?.parsed;
// parsed is typed as { name: string; authors: string[] }

Why good: Schema passed directly to responseFormat, message.parsed is fully typed, named constants

See: examples/structured-output.md for JSON mode, complex schemas

Pattern 5: Function Calling / Tool Use

Define tools and handle the tool call loop.

const tools = [
  {
    type: "function" as const,
    function: {
      name: "get_weather",
      description: "Get current weather for a city",
      parameters: {
        type: "object",
        properties: {
          location: { type: "string", description: "City name" },
        },
        required: ["location"],
      },
    },
  },
];

const result = await client.chat.complete({
  model: "mistral-large-latest",
  messages: [{ role: "user", content: "Weather in Paris?" }],
  tools,
  toolChoice: "any",
});

const toolCall = result?.choices?.[0]?.message?.toolCalls?.[0];
if (toolCall) {
  const args = JSON.parse(toolCall.function.arguments);
  console.log(`Call ${toolCall.function.name} with:`, args);
}

Why good: Uses toolChoice (camelCase), toolCalls (camelCase), proper as const for type literal

See: examples/function-calling.md for complete tool loop, parallel calls

Pattern 6: Embeddings

Create embeddings with mistral-embed. Note: uses inputs (plural), not input.

const EMBEDDING_MODEL = "mistral-embed";

const result = await client.embeddings.create({
  model: EMBEDDING_MODEL,
  inputs: ["First document", "Second document", "Third document"],
});

const vectors = result.data?.map((item) => item.embedding) ?? [];

Why good: Uses inputs (Mistral-specific, plural), named model constant, safe optional chaining

// BAD: Using singular 'input' (OpenAI pattern)
const result = await client.embeddings.create({
  model: "mistral-embed",
  input: ["First document"], // WRONG: Mistral uses 'inputs' (plural)
});

Why bad: Mistral SDK uses inputs (plural) -- input (singular) will error or be silently ignored

See: examples/embeddings-vision.md for cosine similarity, semantic search

Pattern 7: Vision

Send images to vision-capable models using multi-part content arrays.

const result = await client.chat.complete({
  model: "mistral-small-latest",
  messages: [
    {
      role: "user",
      content: [
        { type: "text", text: "What is in this image?" },
        {
          type: "image_url",
          imageUrl: "https://example.com/photo.jpg",
        },
      ],
    },
  ],
});

Why good: Uses imageUrl (camelCase string), not image_url: { url } (OpenAI's nested object pattern)

See: examples/embeddings-vision.md for base64 images, multiple images

Pattern 8: Codestral FIM

Fill-in-middle code completion using the dedicated FIM endpoint.

const result = await client.fim.complete({
  model: "codestral-latest",
  prompt: "function fibonacci(n: number): number {\n  if (n <= 1) return n;\n",
  suffix: "}\n\nconsole.log(fibonacci(10));",
  temperature: 0,
});

const completion = result.choices?.[0]?.message?.content;
// completion fills the gap between prompt and suffix

Why good: Dedicated FIM endpoint, separate prompt + suffix (not messages), deterministic with temperature: 0

See: examples/codestral.md for code generation patterns

Pattern 9: Error Handling

The SDK provides specific error types. Configure retries since the default is no retries.

import { Mistral } from "@mistralai/mistralai";
import {
  SDKError,
  SDKValidationError,
  HTTPValidationError,
} from "@mistralai/mistralai/models/errors";

try {
  const result = await client.chat.complete({
    model: "mistral-large-latest",
    messages: [{ role: "user", content: "Hello" }],
  });
} catch (error) {
  if (error instanceof HTTPValidationError) {
    console.error("Validation error:", error.message);
  } else if (error instanceof SDKValidationError) {
    console.error("Input validation error:", error.message);
  } else if (error instanceof SDKError) {
    console.error(`API error [${error.statusCode}]: ${error.message}`);
  } else {
    throw error;
  }
}

Why good: Specific error types checked in order of specificity, re-throws unexpected errors

See: examples/core.md for full error handling, timeout handling, retry configuration

Performance Optimization

Model Selection

General purpose (most capable)    -> mistral-large-latest (Mistral Large 3)
Balanced cost/quality             -> mistral-medium-latest (Mistral Medium 3.1)
Cost-sensitive / fast             -> mistral-small-latest (Mistral Small 4)
Edge / minimal                    -> ministral-3b-latest or ministral-8b-latest
Complex reasoning                 -> magistral-medium-latest
Code generation                   -> codestral-latest or devstral-latest
Code completion (FIM)             -> codestral-latest (dedicated FIM endpoint)
Vision / images                   -> mistral-small-latest or mistral-large-latest
Embeddings                        -> mistral-embed (1024 dimensions)
Code embeddings                   -> codestral-embed-latest

Key Optimization Patterns

• Configure retries -- Default is no retries. Always set retryConfig for production.
• Set timeouts -- Default is no timeout (-1). Set timeoutMs to avoid hanging requests.
• Use temperature: 0 for deterministic output (enables server-side caching).
• Batch embedding inputs -- Pass multiple strings to inputs array in one call.
• Use FIM for code completion -- fim.complete() is purpose-built and more efficient than chat for code completion tasks.

<decision_framework>

Decision Framework

Which Method to Use

What do you need?
+-- Chat completion (text in, text out)?
|   +-- Need streaming? -> client.chat.stream()
|   +-- Need structured JSON? -> client.chat.parse() with Zod schema
|   +-- Basic completion? -> client.chat.complete()
+-- Code completion / fill-in-middle?
|   +-- YES -> client.fim.complete() with Codestral
+-- Embeddings for search/RAG?
|   +-- YES -> client.embeddings.create() with mistral-embed
+-- Pre-configured agent?
    +-- YES -> client.agents.complete() with agent ID

Which Model to Choose

What is your task?
+-- Most capable general purpose -> mistral-large-latest
+-- Balanced cost/performance -> mistral-medium-latest
+-- Fast + cost-efficient -> mistral-small-latest
+-- Minimal / edge deployment -> ministral-3b-latest
+-- Complex reasoning / math -> magistral-medium-latest
+-- Code generation (chat) -> codestral-latest or devstral-latest
+-- Code completion (FIM) -> codestral-latest
+-- Vision / image analysis -> mistral-small-latest (or any vision-capable model)
+-- Embeddings -> mistral-embed
+-- Code embeddings -> codestral-embed-latest

Streaming vs Non-Streaming

Is the response user-facing?
+-- YES -> Use client.chat.stream()
|   +-- Iterate with: for await (const event of result)
|   +-- Access content: event.data.choices[0]?.delta?.content
+-- NO -> Use client.chat.complete()
    +-- Background processing -> chat.complete()
    +-- Structured output -> chat.parse() with Zod

When to Use This SDK vs a Provider-Agnostic SDK

Do you need multiple LLM providers (Mistral + others)?
+-- YES -> Not this skill's scope -- use a unified provider SDK
+-- NO -> Do you need Mistral-specific features?
    +-- YES -> Use Mistral SDK directly
    |   Examples: Codestral FIM, Mistral Agents,
    |   Voxtral audio, OCR, custom endpoints
    +-- NO -> Mistral SDK is simplest for Mistral-only use

</decision_framework>

<red_flags>

RED FLAGS

High Priority Issues:

• Using response_format (snake_case) instead of responseFormat (camelCase) -- silently ignored, no error thrown
• Using input (singular) for embeddings instead of inputs (plural) -- Mistral-specific naming
• Not configuring retries for production (SDK defaults to strategy: "none" -- zero retries)
• Hardcoding API keys instead of using environment variables
• Accessing chunk.choices[0]?.delta?.content directly on streaming events instead of event.data.choices[0]?.delta?.content

Medium Priority Issues:

• Not setting timeoutMs for production (default is -1, meaning no timeout -- requests can hang indefinitely)
• Using max_tokens instead of maxTokens (camelCase SDK convention)
• Missing system role message for behavior guidance
• Using tool_choice instead of toolChoice
• Using tool_calls instead of toolCalls when reading responses

Common Mistakes:

• Importing from "mistralai" instead of "@mistralai/mistralai" -- the correct package name has the org scope
• Using CommonJS require() -- the package is ESM-only, use import or await import()
• Confusing Mistral's imageUrl: "url" (flat string) with OpenAI's image_url: { url: "..." } (nested object)
• Using client.chat.completions.create() (OpenAI pattern) instead of client.chat.complete() (Mistral pattern)
• Assuming embedding dimensions match OpenAI's -- mistral-embed returns 1024-dimensional vectors, not 1536

Gotchas & Edge Cases:

• The SDK is ESM-only. In CommonJS projects, you must use const { Mistral } = await import("@mistralai/mistralai").
• Streaming content may be string | string[] -- cast or check type when writing to stdout.
• chat.parse() requires a Zod schema passed to responseFormat -- it does not accept { type: "json_object" }.
• The apiKey constructor option accepts a string OR an async function () => Promise<string> for dynamic key rotation.
• Model aliases like mistral-large-latest resolve to the latest version of that model tier. Pin to specific versions (e.g., mistral-large-3-25-12) for reproducibility.
• toolChoice: "any" forces the model to call a tool. toolChoice: "auto" lets the model decide. toolChoice: "none" prevents tool calls.
• parallelToolCalls: false forces sequential tool calling (default true allows parallel).
• FIM endpoint (fim.complete()) uses prompt + suffix parameters, NOT the messages array.
• safePrompt: true injects Mistral's safety system prompt before your messages.
• The SDK provides standalone functions (e.g., chatComplete() from "@mistralai/mistralai/funcs/chatComplete.js") for tree-shaking in browser/edge runtimes.

</red_flags>

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use responseFormat (camelCase) in SDK calls -- NOT response_format (snake_case). The SDK uses camelCase property names throughout.)

(You MUST configure retries explicitly -- the SDK defaults to strategy: "none" (no retries), unlike OpenAI's SDK which retries automatically)

(You MUST consume streaming results with for await (const event of result) and access content via event.data.choices[0]?.delta?.content -- the event shape differs from OpenAI)

(You MUST never hardcode API keys -- use process.env["MISTRAL_API_KEY"] with the bracket notation the SDK documents)

(You MUST use client.chat.parse() with a Zod schema for structured outputs -- NOT manual JSON.parse() on completion content)

Failure to follow these rules will produce broken API calls (snake_case properties silently ignored), unreliable production services (no retries), or incorrectly parsed streaming data.

</critical_reminders>