feat(client): add Realtime API support (#1266)

README.md

@@ -83,6 +83,93 @@ main();
 If you need to cancel a stream, you can `break` from the loop
 or call `stream.controller.abort()`.
 
+## Realtime API beta
+
+The Realtime API enables you to build low-latency, multi-modal conversational experiences. It currently supports text and audio as both input and output, as well as [function calling](https://platform.openai.com/docs/guides/function-calling) through a `WebSocket` connection.
+
+The Realtime API works through a combination of client-sent events and server-sent events. Clients can send events to do things like update session configuration or send text and audio inputs. Server events confirm when audio responses have completed, or when a text response from the model has been received. A full event reference can be found [here](https://platform.openai.com/docs/api-reference/realtime-client-events) and a guide can be found [here](https://platform.openai.com/docs/guides/realtime).
+
+This SDK supports accessing the Realtime API through the [WebSocket API](https://developer.mozilla.org/en-US/docs/Web/API/WebSocket) or with [ws](https://github.com/websockets/ws).
+
+Basic text based example with `ws`:
+
+```ts
+// requires `yarn add ws @types/ws`
+import { OpenAIRealtimeWS } from 'openai/beta/realtime/ws';
+
+const rt = new OpenAIRealtimeWS({ model: 'gpt-4o-realtime-preview-2024-12-17' });
+
+// access the underlying `ws.WebSocket` instance
+rt.socket.on('open', () => {
+  console.log('Connection opened!');
+  rt.send({
+    type: 'session.update',
+    session: {
+      modalities: ['text'],
+      model: 'gpt-4o-realtime-preview',
+    },
+  });
+
+  rt.send({
+    type: 'conversation.item.create',
+    item: {
+      type: 'message',
+      role: 'user',
+      content: [{ type: 'input_text', text: 'Say a couple paragraphs!' }],
+    },
+  });
+
+  rt.send({ type: 'response.create' });
+});
+
+rt.on('error', (err) => {
+  // in a real world scenario this should be logged somewhere as you
+  // likely want to continue procesing events regardless of any errors
+  throw err;
+});
+
+rt.on('session.created', (event) => {
+  console.log('session created!', event.session);
+  console.log();
+});
+
+rt.on('response.text.delta', (event) => process.stdout.write(event.delta));
+rt.on('response.text.done', () => console.log());
+
+rt.on('response.done', () => rt.close());
+
+rt.socket.on('close', () => console.log('\nConnection closed!'));
+```
+
+To use the web API `WebSocket` implementation, replace `OpenAIRealtimeWS` with `OpenAIRealtimeWebSocket` and adjust any `rt.socket` access:
+
+```ts
+import { OpenAIRealtimeWebSocket } from 'openai/beta/realtime/websocket';
+
+const rt = new OpenAIRealtimeWebSocket({ model: 'gpt-4o-realtime-preview-2024-12-17' });
+// ...
+rt.socket.addEventListener('open', () => {
+ // ...
+});
+```
+
+A full example can be found [here](https://github.com/openai/openai-node/blob/master/examples/realtime/web.ts).
+
+### Realtime error handling
+
+When an error is encountered, either on the client side or returned from the server through the [`error` event](https://platform.openai.com/docs/guides/realtime/realtime-api-beta#handling-errors), the `error` event listener will be fired. However, if you haven't registered an `error` event listener then an `unhandled Promise rejection` error will be thrown.
+
+It is **highly recommended** that you register an `error` event listener and handle errors approriately as typically the underlying connection is still usable.
+
+```ts
+const rt = new OpenAIRealtimeWS({ model: 'gpt-4o-realtime-preview-2024-12-17' });
+rt.on('error', (err) => {
+  // in a real world scenario this should be logged somewhere as you
+  // likely want to continue procesing events regardless of any errors
+  throw err;
+});
+```
+
 ### Request & Response types
 
 This library includes TypeScript definitions for all request params and response fields. You may import and use them like so:

examples/package.json

@@ -6,14 +6,15 @@
   "license": "MIT",
   "private": true,
   "dependencies": {
+    "@azure/identity": "^4.2.0",
     "express": "^4.18.2",
     "next": "^14.1.1",
     "openai": "file:..",
-    "zod-to-json-schema": "^3.21.4",
-    "@azure/identity": "^4.2.0"
+    "zod-to-json-schema": "^3.21.4"
   },
   "devDependencies": {
     "@types/body-parser": "^1.19.3",
-    "@types/express": "^4.17.19"
+    "@types/express": "^4.17.19",
+    "@types/web": "^0.0.194"
   }
 }

examples/realtime/websocket.ts

@@ -0,0 +1,48 @@
+import { OpenAIRealtimeWebSocket } from 'openai/beta/realtime/websocket';
+
+async function main() {
+  const rt = new OpenAIRealtimeWebSocket({ model: 'gpt-4o-realtime-preview-2024-12-17' });
+
+  // access the underlying `ws.WebSocket` instance
+  rt.socket.addEventListener('open', () => {
+    console.log('Connection opened!');
+    rt.send({
+      type: 'session.update',
+      session: {
+        modalities: ['text'],
+        model: 'gpt-4o-realtime-preview',
+      },
+    });
+
+    rt.send({
+      type: 'conversation.item.create',
+      item: {
+        type: 'message',
+        role: 'user',
+        content: [{ type: 'input_text', text: 'Say a couple paragraphs!' }],
+      },
+    });
+
+    rt.send({ type: 'response.create' });
+  });
+
+  rt.on('error', (err) => {
+    // in a real world scenario this should be logged somewhere as you
+    // likely want to continue procesing events regardless of any errors
+    throw err;
+  });
+
+  rt.on('session.created', (event) => {
+    console.log('session created!', event.session);
+    console.log();
+  });
+
+  rt.on('response.text.delta', (event) => process.stdout.write(event.delta));
+  rt.on('response.text.done', () => console.log());
+
+  rt.on('response.done', () => rt.close());
+
+  rt.socket.addEventListener('close', () => console.log('\nConnection closed!'));
+}
+
+main();

examples/realtime/ws.ts

@@ -0,0 +1,55 @@
+import { OpenAIRealtimeWS } from 'openai/beta/realtime/ws';
+
+async function main() {
+  const rt = new OpenAIRealtimeWS({ model: 'gpt-4o-realtime-preview-2024-12-17' });
+
+  // access the underlying `ws.WebSocket` instance
+  rt.socket.on('open', () => {
+    console.log('Connection opened!');
+    rt.send({
+      type: 'session.update',
+      session: {
+        modalities: ['foo'] as any,
+        model: 'gpt-4o-realtime-preview',
+      },
+    });
+    rt.send({
+      type: 'session.update',
+      session: {
+        modalities: ['text'],
+        model: 'gpt-4o-realtime-preview',
+      },
+    });
+
+    rt.send({
+      type: 'conversation.item.create',
+      item: {
+        type: 'message',
+        role: 'user',
+        content: [{ type: 'input_text', text: 'Say a couple paragraphs!' }],
+      },
+    });
+
+    rt.send({ type: 'response.create' });
+  });
+
+  rt.on('error', (err) => {
+    // in a real world scenario this should be logged somewhere as you
+    // likely want to continue procesing events regardless of any errors
+    throw err;
+  });
+
+  rt.on('session.created', (event) => {
+    console.log('session created!', event.session);
+    console.log();
+  });
+
+  rt.on('response.text.delta', (event) => process.stdout.write(event.delta));
+  rt.on('response.text.done', () => console.log());
+
+  rt.on('response.done', () => rt.close());
+
+  rt.socket.on('close', () => console.log('\nConnection closed!'));
+}
+
+main();

package.json

@@ -36,6 +36,7 @@
     "@swc/core": "^1.3.102",
     "@swc/jest": "^0.2.29",
     "@types/jest": "^29.4.0",
+    "@types/ws": "^8.5.13",
     "@typescript-eslint/eslint-plugin": "^6.7.0",
     "@typescript-eslint/parser": "^6.7.0",
     "eslint": "^8.49.0",
@@ -52,6 +53,7 @@
     "tsc-multi": "^1.1.0",
     "tsconfig-paths": "^4.0.0",
     "typescript": "^4.8.2",
+    "ws": "^8.18.0",
     "zod": "^3.23.8"
   },
   "sideEffects": [
@@ -126,9 +128,13 @@
   },
   "bin": "./bin/cli",
   "peerDependencies": {
+    "ws": "^8.18.0",
     "zod": "^3.23.8"
   },
   "peerDependenciesMeta": {
+    "ws": {
+      "optional": true
+    },
     "zod": {
       "optional": true
     }

src/beta/realtime/index.ts

@@ -0,0 +1 @@
+export { OpenAIRealtimeError } from './internal-base';

src/beta/realtime/internal-base.ts

@@ -0,0 +1,83 @@
+import { RealtimeClientEvent, RealtimeServerEvent, ErrorEvent } from '../../resources/beta/realtime/realtime';
+import { EventEmitter } from '../../lib/EventEmitter';
+import { OpenAIError } from '../../error';
+
+export class OpenAIRealtimeError extends OpenAIError {
+  /**
+   * The error data that the API sent back in an `error` event.
+   */
+  error?: ErrorEvent.Error | undefined;
+
+  /**
+   * The unique ID of the server event.
+   */
+  event_id?: string | undefined;
+
+  constructor(message: string, event: ErrorEvent | null) {
+    super(message);
+
+    this.error = event?.error;
+    this.event_id = event?.event_id;
+  }
+}
+
+type Simplify<T> = { [KeyType in keyof T]: T[KeyType] } & {};
+
+type RealtimeEvents = Simplify<
+  {
+    event: (event: RealtimeServerEvent) => void;
+    error: (error: OpenAIRealtimeError) => void;
+  } & {
+    [EventType in Exclude<RealtimeServerEvent['type'], 'error'>]: (
+      event: Extract<RealtimeServerEvent, { type: EventType }>,
+    ) => unknown;
+  }
+>;
+
+export abstract class OpenAIRealtimeEmitter extends EventEmitter<RealtimeEvents> {
+  /**
+   * Send an event to the API.
+   */
+  abstract send(event: RealtimeClientEvent): void;
+
+  /**
+   * Close the websocket connection.
+   */
+  abstract close(props?: { code: number; reason: string }): void;
+
+  protected _onError(event: null, message: string, cause: any): void;
+  protected _onError(event: ErrorEvent, message?: string | undefined): void;
+  protected _onError(event: ErrorEvent | null, message?: string | undefined, cause?: any): void {
+    message =
+      event?.error ?
+        `${event.error.message} code=${event.error.code} param=${event.error.param} type=${event.error.type} event_id=${event.error.event_id}`
+      : message ?? 'unknown error';
+
+    if (!this._hasListener('error')) {
+      const error = new OpenAIRealtimeError(
+        message +
+          `\n\nTo resolve these unhandled rejection errors you should bind an \`error\` callback, e.g. \`rt.on('error', (error) => ...)\` `,
+        event,
+      );
+      // @ts-ignore
+      error.cause = cause;
+      Promise.reject(error);
+      return;
+    }
+
+    const error = new OpenAIRealtimeError(message, event);
+    // @ts-ignore
+    error.cause = cause;
+
+    this._emit('error', error);
+  }
+}
+
+export function buildRealtimeURL(props: { baseURL: string; model: string }): URL {
+  const path = '/realtime';
+
+  const url = new URL(props.baseURL + (props.baseURL.endsWith('/') ? path.slice(1) : path));
+  url.protocol = 'wss';
+  url.searchParams.set('model', props.model);
+  return url;
+}