@braintrust/lingua

TypeScript bindings for Lingua - a universal message format for LLMs.

Installation

pnpm add @braintrust/lingua
# or
npm install @braintrust/lingua
# or
yarn add @braintrust/lingua

This automatically installs @braintrust/lingua-wasm as a dependency (you don't need to install it separately).

If you only need TypeScript types and do not need conversion or validation functions, install the types-only package instead:

pnpm add -D @braintrust/lingua-types

import type { Message, UniversalRequest } from "@braintrust/lingua-types";

Usage

The package provides separate entry points for Node.js and browser environments:

• @braintrust/lingua or @braintrust/lingua/node - Node.js with native WASM (auto-initialized)
• @braintrust/lingua/browser - Browser with web-targeted WASM (requires explicit initialization)

Node.js

In Node.js, the WASM module is automatically loaded - just import and use:

import {
  type Message,
  linguaToChatCompletionsMessages,
  chatCompletionsMessagesToLingua,
} from "@braintrust/lingua";

// Create messages in Lingua format
const messages: Message[] = [
  { role: "user", content: "Hello, how are you?" },
];

// Convert to OpenAI format
const openaiMessages = linguaToChatCompletionsMessages(messages);

// Convert OpenAI response back to Lingua
const linguaMessages = chatCompletionsMessagesToLingua([response.choices[0].message]);

Browser

The browser build requires explicit initialization before use:

import init, {
  linguaToChatCompletionsMessages,
  type Message,
} from "@braintrust/lingua/browser";

// Initialize with WASM URL (must be called before using any functions)
await init("/wasm/lingua.wasm");

// Now you can use Lingua functions
const messages: Message[] = [{ role: "user", content: "Hello!" }];
const openaiMessages = linguaToChatCompletionsMessages(messages);

Initialization options:

// Option 1: URL string (fetches the WASM file)
await init("/wasm/lingua.wasm");

// Option 2: ArrayBuffer/Uint8Array (useful for testing or custom loading)
const wasmBuffer = await fetch("/wasm/lingua.wasm").then((r) => r.arrayBuffer());
await init(wasmBuffer);

Important: You must provide a WASM source to init(). The function will throw an error if called without an argument.

Next.js

For Next.js applications, you need to:

1. Copy the WASM file to your public directory during build
2. Exclude the WASM package from webpack processing
3. Initialize on the client side

Step 1: Configure next.config.mjs

import CopyWebpackPlugin from "copy-webpack-plugin";

/** @type {import('next').NextConfig} */
const nextConfig = {
  webpack: (config, { isServer }) => {
    // Copy WASM file to public directory
    config.plugins.push(
      new CopyWebpackPlugin({
        patterns: [
          {
            from: "node_modules/@braintrust/lingua-wasm/web/lingua_bg.wasm",
            to: "../public/wasm/lingua.wasm",
          },
        ],
      })
    );

    // Exclude lingua-wasm from webpack's WASM processing
    config.module.rules.push({
      test: /\.wasm$/,
      exclude: [/[\\/]lingua-wasm[\\/]/],
      type: "webassembly/async",
    });

    return config;
  },
  // Mark the WASM package as external for server-side
  serverExternalPackages: ["@braintrust/lingua-wasm"],
};

export default nextConfig;

Step 2: Create a client-side wrapper

// lib/lingua.tsx
"use client";

import * as linguaModule from "@braintrust/lingua/browser";
import { useEffect } from "react";

const WASM_URL = "/wasm/lingua.wasm";

let ready: Promise<void> | null = null;

export async function initLingua(): Promise<void> {
  if (!ready) {
    ready = linguaModule.init(WASM_URL).catch((error) => {
      console.error("[Lingua] Initialization failed:", error);
      ready = null;
      throw error;
    });
  }
  return ready;
}

// Optional: React provider to initialize on app load
export function LinguaProvider({ children }: { children: React.ReactNode }) {
  useEffect(() => {
    initLingua();
  }, []);

  return <>{children}</>;
}

// Re-export everything from the browser module
export * from "@braintrust/lingua/browser";

Step 3: Use in your components

"use client";

import { initLingua, linguaToChatCompletionsMessages } from "@/lib/lingua";

async function convertMessages() {
  await initLingua(); // Ensures WASM is loaded

  const messages = [{ role: "user" as const, content: "Hello!" }];
  return linguaToChatCompletionsMessages(messages);
}

Package architecture

The TypeScript packages are split by runtime needs:

@braintrust/lingua          # TypeScript wrapper (this package)
  └── @braintrust/lingua-wasm  # Raw WASM bindings (auto-installed dependency)
        ├── nodejs/            # Node.js WASM build
        └── web/               # Browser WASM build
@braintrust/lingua-types    # Type-only package with no WASM dependency

• @braintrust/lingua - Pure TypeScript that imports from @braintrust/lingua-wasm
• @braintrust/lingua-wasm - Raw wasm-pack output, separate package for clean bundling
• @braintrust/lingua-types - Bundled TypeScript declarations for consumers that only need types

This separation ensures webpack/bundlers can properly handle the WASM files without complex configuration.

Development

This package is part of the Lingua monorepo. The TypeScript types are automatically generated from the Rust source code using ts-rs.

Building

# Build TypeScript types, WASM, and wrapper packages from the repo root
make typescript

Generating Types

To regenerate the TypeScript types from Rust:

pnpm generate

Running Tests

pnpm test

Project Structure

bindings/
├── lingua-wasm/           # @braintrust/lingua-wasm package
│   ├── nodejs/            # wasm-pack --target nodejs output
│   ├── web/               # wasm-pack --target web output
│   └── package.json
├── typescript-types/      # @braintrust/lingua-types package
│   ├── src/
│   └── dist/
└── typescript/            # @braintrust/lingua package (this directory)
    ├── src/
    │   ├── generated/         # Auto-generated types from Rust
    │   ├── index.ts           # Node.js entry point
    │   ├── index.browser.ts   # Browser entry point
    │   └── wasm-runtime.ts    # WASM initialization logic
    ├── dist/
    │   ├── index.mjs          # Compiled Node.js module
    │   ├── index.browser.mjs  # Compiled browser module
    │   └── *.d.mts            # Type definitions
    ├── tests/
    └── package.json

API Reference

Conversion Functions

// OpenAI Chat Completions ↔ Lingua
linguaToChatCompletionsMessages(messages: Message[]): ChatCompletionMessageParam[]
chatCompletionsMessagesToLingua(messages: ChatCompletionMessage[]): Message[]

// Anthropic ↔ Lingua
linguaToAnthropicMessages(messages: Message[]): MessageParam[]
anthropicMessagesToLingua(messages: AnthropicMessage[]): Message[]

// Validation
validateChatCompletionsRequest(request: unknown): ValidationResult
validateChatCompletionsResponse(response: unknown): ValidationResult
validateAnthropicRequest(request: unknown): ValidationResult
validateAnthropicResponse(response: unknown): ValidationResult

Types

interface Message {
  role: "user" | "assistant" | "system" | "tool";
  content: string | ContentPart[];
  id?: string | null;
}

Type Compatibility

The generated types are designed to be compatible with popular LLM SDKs:

• OpenAI SDK (openai)
• Anthropic SDK (@anthropic-ai/sdk)

License

MIT