Enabling Event-Streaming Operations

Server-sent events (SSE) is a core web feature that provides servers with a low overhead solution to push real-time events to the client when they become available. SSE can be used to stream chat completions from a large language model, real-time stock prices, and sensor readings to clients.

SSE is similar to WebSockets in that it uses a persistent connection but differs in that it is unidirectional - only the server sends events. SSE is simpler to implement in many existing backend HTTP frameworks.

INFO

Speakeasy makes it easy to build SSE into your SDKs without vendor extensions or heuristics. Leverage SSE by modeling SSE streams as text/event-stream responses with pure OpenAPI.

Here’s a short example of using an SDK to chat with an LLM and read its response as a stream:

import { SDK } from '@speakeasy/sdk';

const sdk = new SDK()

const response = await sdk.chat.create({
  prompt: "What are the top 3 French cheeses by consumption?"
})

for await (const event of response.chatStream) {
  process.stdout.write(event.data);
}

INFO

The SSE feature is currently supported in TypeScript, Python, Go, and Java. Let us know if you’d like to see support for other languages.

Modeling SSE in OpenAPI

To implement SSE in your SDKs, model an API endpoint that serves an event stream in your OpenAPI document. Each server-sent event can contain up to four types of fields: id, event, data, and retry.

The example below illustrates an operation that streams events containing only a data field that holds string content:

paths:
  /chat:
    post:
      summary: Create a chat completion from a prompt
      operationId: create
      tags: [chat]
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/ChatRequest'
      responses:
        '200':
          description: Chat completion created
          content:
            text/event-stream:
              schema:
                $ref: '#/components/schemas/ChatStream'
components:
  schemas:
    ChatRequest:
      type: object
      required: [prompt]
      properties:
        prompt:
          type: string
    ChatStream:
      description: A server-sent event containing chat completion content
      type: object
      required: [data]
      properties:
        data:
          type: string

SSE implementation isn’t limited to string data. If you specify that data is an object, then SDKs will assume the field will contain JSON content. Raw data received from the server will be deserialized into an object for the application code to consume.

components:
  schemas:
    ChatStream:
      description: A server-sent event containing chat completion content
      type: object
      required: [data]
      properties:
        data:
          type: object
          properties:
            content:
              type: string
            model:
              type: string
              enum: ["foo-gpt-tiny", "foo-gpt-small"]
            created:
              type: integer

The Speakeasy-generated TypeScript SDK for the example above will allow users to access this object:

for await (const event of response.chatStream) {
  const { content, model, created } = event.data;

  process.stdout.write(content);
}

Other streaming APIs send multiple types of events with the id and event fields. These event types can be described as a union (oneOf) with the event field acting as a discriminator:

components:
  schemas:
    ChatStream:
      oneOf:
        - $ref: '#/components/schemas/HeartbeatEvent'
        - $ref: '#/components/schemas/ChatEvent'
      discriminator:
        propertyName: event
        mapping:
          ping: '#/components/schemas/HeartbeatEvent'
          completion: '#/components/schemas/ChatEvent'

    HeartbeatEvent:
      description: A server-sent event indicating that the server is still processing the request
      type: object
      required: [event]
      properties:
        event:
          type: string
          const: "ping"
    ChatEvent:
      description: A server-sent event containing chat completion content
      type: object
      required: [id, event, data]
      properties:
        id:
          type: string
        event:
          type: string
          const: completion
        data:
          type: object
          required: [content]
          properties:
            content:
              type: string

Note that across all these examples, the schema for the events only ever specifies one or more of the four recognized fields. Adding other fields will trigger a validation error when generating an SDK with the Speakeasy CLI or GitHub action.

Sentinel events

Some SSE APIs will terminate the stream by sending a final, special event. This sentinel event is only used to signal that there are no more events and is not intended for application code to handle.

In the example below, the final data: [DONE] event is the sentinel event:

HTTP/1.1 200 OK
Content-Type: text/event-stream; charset=utf-8
Date: Fri, 12 Jul 2024 14:29:22 GMT
Keep-Alive: timeout=5, max=1000
Connection: Keep-Alive

data: {"content": "there"}

data: {"content": "are 7"}

data: {"content": "continents in the world"}

data: [DONE]

To hide this final event in generated SDK methods, use the x-speakeasy-sse-sentinel: <string> extension on a text/event-stream media object:

paths:
  /chat:
    post:
      summary: Create a chat completion from a prompt
      operationId: create
      tags: [chat]
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/ChatRequest'
      responses:
        '200':
          description: Chat completion created
          content:
            text/event-stream:
+             x-speakeasy-sse-sentinel: '[DONE]'
              schema:
                $ref: '#/components/schemas/ChatEvent'

components:
  schemas:
    ChatEvent:
      description: A server-sent event containing chat completion content
      type: object
      required: [data]
      properties:
        data:
          type: object
          required: [content]
          properties:
            content:
              type: string

Application code like the following TypeScript sample will behave as expected. The async iteration loop will finish when the sentinel event is encountered:

const llm = new LLM();

const stream = await llm.chat.create({
  prompt: "How many continents are there?",
});

for await (const event of stream) {
//               ^? ChatEvent
  process.stdout.write(event.data.content);
}