examples/openai-realtime/README.md
This example demonstrates how to use promptfoo to test OpenAI's Realtime API capabilities. The Realtime API allows for real-time communication with gpt-realtime-2, gpt-realtime-1.5, and gpt-realtime using WebSockets, supporting text, audio, and model-dependent image inputs plus text/audio outputs.
You can run this example with:
npx promptfoo@latest init --example openai-realtime
cd openai-realtime
This will create all necessary files and folder structure to get started quickly.
export OPENAI_API_KEY=your-api-key-here
You can point the Realtime provider at custom/proxy endpoints (including Azure-compatible gateways) or local/dev servers by setting apiBaseUrl. The provider automatically converts https:// → wss:// and http:// → ws:// for the WebSocket connection.
providers:
- id: openai:realtime:gpt-realtime-1.5
config:
# Custom hosted gateway
apiBaseUrl: 'https://my-custom-api.com/v1' # connects to wss://my-custom-api.com/v1/realtime
modalities: ['text']
For local development:
providers:
- id: openai:realtime:gpt-realtime-1.5
config:
apiBaseUrl: 'http://localhost:8080/v1' # connects to ws://localhost:8080/v1/realtime
modalities: ['text']
You can also use environment variables like OPENAI_API_BASE_URL or OPENAI_BASE_URL instead of apiBaseUrl.
promptfooconfig.yaml: Configuration file defining the providers and testspromptfooconfig-gpt-realtime.yaml: Comprehensive gpt-realtime-1.5 model demonstration with audio supportpromptfooconfig-function-calling.js: Runnable function-calling example with a local handlertest-webui-audio.yaml: Simple audio test for WebUI playbackrealtime-input.json: JSON template for the realtime input promptpromptfooconfig-conversation.yaml: Configuration for multi-turn conversation testsrealtime-conversation.js: JavaScript prompt function for multi-turn conversationsThe Realtime API supports maintaining conversation history across multiple turns. This example includes a multi-turn conversation configuration that demonstrates how to:
To run the multi-turn conversation example:
npx promptfoo eval -c examples/openai-realtime/promptfooconfig-conversation.yaml
The multi-turn conversation example demonstrates how the OpenAI Realtime API can maintain context across multiple exchanges. This is implemented using promptfoo's built-in support for conversation history through the _conversation variable and metadata.
_conversation variable contains all previous turns in the conversationconversationId metadata value is part of the same conversation threadWhen using conversationId in the metadata of tests, promptfoo automatically:
For each conversation turn:
_conversation variable is automatically populated with all previous prompts and outputsUser: What are some popular tourist destinations in Japan?
AI: Some popular tourist destinations in Japan include Tokyo, Kyoto, Osaka, Hiroshima, and Hokkaido...
User: Which of those places is best to visit in autumn?
AI: Kyoto is particularly beautiful in autumn with its colorful maple leaves...
User: What traditional foods should I try there?
AI: In Kyoto during autumn, you should try momiji manju (maple leaf-shaped cakes), kyo-kaiseki (traditional multi-course meal)...
The API maintains context throughout this exchange, understanding that follow-up questions refer to Japan and then to the specific autumn locations.
This example uses a JavaScript function (realtime-conversation.js) to properly format the conversation for the OpenAI Realtime API:
module.exports = async function ({ vars, provider }) {
// Create the messages array starting with system message
const messages = [
{
role: 'system',
content: [
{
type: 'input_text',
text: vars.system_message || 'You are a helpful AI assistant.',
},
],
},
];
// Add previous conversation turns if they exist
if (vars._conversation && Array.isArray(vars._conversation)) {
for (const completion of vars._conversation) {
// Add user message
messages.push({
role: 'user',
content: [
{
type: 'input_text',
text: completion.input,
},
],
});
// Add assistant message
messages.push({
role: 'assistant',
content: [
{
type: 'text',
text: completion.output,
},
],
});
}
}
// Add the current question as the final user message
messages.push({
role: 'user',
content: [
{
type: 'input_text',
text: vars.question || '',
},
],
});
return messages;
};
This approach provides better flexibility and error handling than using JSON templates with Nunjucks.
We also provide a JSON template approach for reference:
[
{
"role": "system",
"content": [
{
"type": "input_text",
"text": "{{ system_message }}"
}
]
}{% for completion in _conversation %},
{
"role": "user",
"content": [
{
"type": "input_text",
"text": "{{ completion.input }}"
}
]
},
{
"role": "assistant",
"content": [
{
"type": "text",
"text": "{{ completion.output }}"
}
]
}{% endfor %},
{
"role": "user",
"content": [
{
"type": "input_text",
"text": "{{ question }}"
}
]
}
]
Note: JSON validators may show errors for this template because of the Nunjucks expressions, but promptfoo will correctly process this file at runtime. This approach uses the
_conversationvariable to maintain conversation history in a way that works with the Realtime API.
The configuration includes two separate conversation threads:
Each thread maintains its own independent context while tests are evaluated.
The provider implementation in promptfoo creates a direct WebSocket connection with the OpenAI Realtime API, following the official protocol:
wss://api.openai.com/v1/realtime?model=MODEL_IDThe WebSocket connection follows the official OpenAI documentation:
const wsUrl = `wss://api.openai.com/v1/realtime?model=${modelName}`;
const ws = new WebSocket(wsUrl, {
headers: {
Authorization: `Bearer ${apiKey}`,
// Other headers...
},
});
When sending messages to the OpenAI Realtime API, you must use the correct content type format:
// When sending user or system messages, use the 'input_text' content type
ws.send(
JSON.stringify({
type: 'conversation.item.create',
item: {
type: 'message',
role: 'user',
content: [
{
type: 'input_text', // Must be 'input_text' for user/system inputs
text: 'Your message here',
},
],
},
}),
);
// When configuring modalities for response settings, use 'text' and 'audio'
const config = {
modalities: ['text', 'audio'],
};
Structured Realtime prompts can also preserve the native multimodal user-content items documented by OpenAI:
[
{
"role": "user",
"content": [
{
"type": "input_text",
"text": "Describe these inputs."
},
{
"type": "input_audio",
"audio": "<base64-encoded audio>"
},
{
"type": "input_image",
"image_url": "data:image/jpeg;base64,..."
}
]
}
]
Use input_image only with Realtime models that support image input, such as the current gpt-realtime* family.
Promptfoo keeps the existing modalities config key for compatibility, but sends the current GA Realtime wire shape to OpenAI under the hood.
The provider supports the Realtime API's function calling capabilities:
Realtime function tools use the native OpenAI Realtime shape with top-level name, description, and parameters fields:
providers:
- id: openai:realtime:gpt-realtime-1.5
config:
tools:
- type: function
name: get_weather
description: Get the current weather for a location
parameters:
type: object
properties:
location:
type: string
required: ['location']
tool_choice: auto
If you reuse a Chat Completions-style tools file that wraps those fields under function:, promptfoo still accepts it as a compatibility input and normalizes only that legacy shape before sending it to the Realtime API.
When you provide a custom functionCallHandler, promptfoo forwards the model-emitted tool name and arguments to your handler. Use toolCallTimeout to bound each handler invocation and maxToolIterations to cap chained tool-follow-up rounds in one turn. Validate the function name and parse or validate the arguments before side effects in your application code.
To use function calling in your application, implement a function call handler. The runnable promptfooconfig-function-calling.js example includes one end to end; here is the same core pattern:
// In your application code
const functionCallHandler = async (name, args) => {
// Parse the function arguments
const parsedArgs = JSON.parse(args);
if (name === 'get_weather') {
const { location } = parsedArgs;
// In a real implementation, you would call a weather API here
// This is just a mock example
return JSON.stringify({
location,
temperature: '72°F',
condition: 'Sunny',
humidity: '45%',
forecast: 'Clear skies for the next 24 hours',
});
}
// Handle unknown function
return JSON.stringify({ error: `Unknown function: ${name}` });
};
// You can then pass this handler in your prompt configuration
const config = {
functionCallHandler,
};
The Realtime API supports both text and audio interactions. promptfoo now includes full audio support:
cedar and marin (in addition to existing voices)promptfoo view to access the WebUI and play generated audio filesTo enable audio support, configure your provider with:
providers:
- id: openai:realtime:gpt-realtime-1.5
config:
modalities: ['text', 'audio']
voice: 'cedar' # or 'marin', 'alloy', 'ash', 'ballad', 'coral', 'echo', 'sage', 'shimmer', 'verse'
instructions: 'Please respond with audio.'
test-webui-audio.yaml: Simple audio test for WebUI playbackpromptfooconfig-gpt-realtime.yaml: Comprehensive gpt-realtime model demonstrationFrom the root directory of promptfoo, run:
npx promptfoo eval -c examples/openai-realtime/promptfooconfig.yaml
If you encounter a "WebSocket error: Unexpected server response: 403" error, this typically indicates one of these issues:
Network/Firewall Restrictions: WebSocket connections may be blocked by your network or firewall.
API Access: Your OpenAI API key may not have access to the Realtime API.
Rate Limits: You may have hit rate limits or quotas for the Realtime API.
If you're unable to use the Realtime API due to WebSocket connection issues, you can still use the regular OpenAI chat API for most use cases. The configuration includes both providers, so you'll see results from the regular chat API even if the Realtime API fails to connect.