The OpenAI API supports extracting JSON from the model with the response_format
request param, for more details on the API, see this guide.
The SDK provides a client.beta.chat.completions.parse()
method which is a wrapper over the client.chat.completions.create()
that
provides richer integrations with TS specific types & returns a ParsedChatCompletion
object, which is an extension of the standard ChatCompletion
type.
You can pass zod schemas wrapped with zodResponseFormat()
to the .parse()
method and the SDK will automatically conver the model
into a JSON schema, send it to the API and parse the response content back using the given zod schema.
import { zodResponseFormat } from 'openai/helpers/zod';
import OpenAI from 'openai/index';
import { z } from 'zod';
const Step = z.object({
explanation: z.string(),
output: z.string(),
});
const MathResponse = z.object({
steps: z.array(Step),
final_answer: z.string(),
});
const client = new OpenAI();
const completion = await client.beta.chat.completions.parse({
model: 'gpt-4o-2024-08-06',
messages: [
{ role: 'system', content: 'You are a helpful math tutor.' },
{ role: 'user', content: 'solve 8x + 31 = 2' },
],
response_format: zodResponseFormat(MathResponse, 'math_response'),
});
console.dir(completion, { depth: 5 });
const message = completion.choices[0]?.message;
if (message?.parsed) {
console.log(message.parsed.steps);
console.log(`answer: ${message.parsed.final_answer}`);
}
The .parse()
method will also automatically parse function
tool calls if:
- You use the
zodFunctionTool()
helper method - You mark your tool schema with
"strict": True
For example:
import { zodFunction } from 'openai/helpers/zod';
import OpenAI from 'openai/index';
import { z } from 'zod';
const Table = z.enum(['orders', 'customers', 'products']);
const Column = z.enum([
'id',
'status',
'expected_delivery_date',
'delivered_at',
'shipped_at',
'ordered_at',
'canceled_at',
]);
const Operator = z.enum(['=', '>', '<', '<=', '>=', '!=']);
const OrderBy = z.enum(['asc', 'desc']);
const DynamicValue = z.object({
column_name: z.string(),
});
const Condition = z.object({
column: z.string(),
operator: Operator,
value: z.union([z.string(), z.number(), DynamicValue]),
});
const Query = z.object({
table_name: Table,
columns: z.array(Column),
conditions: z.array(Condition),
order_by: OrderBy,
});
const client = new OpenAI();
const completion = await client.beta.chat.completions.parse({
model: 'gpt-4o-2024-08-06',
messages: [
{
role: 'system',
content:
'You are a helpful assistant. The current date is August 6, 2024. You help users query for the data they are looking for by calling the query function.',
},
{
role: 'user',
content: 'look up all my orders in november of last year that were fulfilled but not delivered on time',
},
],
tools: [zodFunction({ name: 'query', parameters: Query })],
});
console.dir(completion, { depth: 10 });
const toolCall = completion.choices[0]?.message.tool_calls?.[0];
if (toolCall) {
const args = toolCall.function.parsed_arguments as z.infer<typeof Query>;
console.log(args);
console.log(args.table_name);
}
main();
The beta.chat.completions.parse()
method imposes some additional restrictions on it's usage that chat.completions.create()
does not.
- If the completion completes with
finish_reason
set tolength
orcontent_filter
, theLengthFinishReasonError
/ContentFilterFinishReasonError
errors will be raised. - Only strict function tools can be passed, e.g.
{type: 'function', function: {..., strict: true}}
OpenAI supports streaming responses when interacting with the Chat or Assistant APIs.
OpenAI supports streaming responses from Assistants. The SDK provides convenience wrappers around the API so you can subscribe to the types of events you are interested in as well as receive accumulated responses.
More information can be found in the documentation: Assistant Streaming
const run = openai.beta.threads.runs
.stream(thread.id, {
assistant_id: assistant.id,
})
.on('textCreated', (text) => process.stdout.write('\nassistant > '))
.on('textDelta', (textDelta, snapshot) => process.stdout.write(textDelta.value))
.on('toolCallCreated', (toolCall) => process.stdout.write(`\nassistant > ${toolCall.type}\n\n`))
.on('toolCallDelta', (toolCallDelta, snapshot) => {
if (toolCallDelta.type === 'code_interpreter') {
if (toolCallDelta.code_interpreter.input) {
process.stdout.write(toolCallDelta.code_interpreter.input);
}
if (toolCallDelta.code_interpreter.outputs) {
process.stdout.write('\noutput >\n');
toolCallDelta.code_interpreter.outputs.forEach((output) => {
if (output.type === 'logs') {
process.stdout.write(`\n${output.logs}\n`);
}
});
}
}
});
There are three helper methods for creating streams:
openai.beta.threads.runs.stream();
This method can be used to start and stream the response to an existing run with an associated thread that is already populated with messages.
openai.beta.threads.createAndRunStream();
This method can be used to add a message to a thread, start a run and then stream the response.
openai.beta.threads.runs.submitToolOutputsStream();
This method can be used to submit a tool output to a run waiting on the output and start a stream.
The assistant API provides events you can subscribe to for the following events.
.on('event', (event: AssistantStreamEvent) => ...)
This allows you to subscribe to all the possible raw events sent by the OpenAI streaming API. In many cases it will be more convenient to subscribe to a more specific set of events for your use case.
More information on the types of events can be found here: Events
.on('runStepCreated', (runStep: RunStep) => ...)
.on('runStepDelta', (delta: RunStepDelta, snapshot: RunStep) => ...)
.on('runStepDone', (runStep: RunStep) => ...)
These events allow you to subscribe to the creation, delta and completion of a RunStep.
For more information on how Runs and RunSteps work see the documentation Runs and RunSteps
.on('messageCreated', (message: Message) => ...)
.on('messageDelta', (delta: MessageDelta, snapshot: Message) => ...)
.on('messageDone', (message: Message) => ...)
This allows you to subscribe to Message creation, delta and completion events. Messages can contain different types of content that can be sent from a model (and events are available for specific content types). For convenience, the delta event includes both the incremental update and an accumulated snapshot of the content.
More information on messages can be found on in the documentation page Message.
.on('textCreated', (content: Text) => ...)
.on('textDelta', (delta: RunStepDelta, snapshot: Text) => ...)
.on('textDone', (content: Text, snapshot: Message) => ...)
These events allow you to subscribe to the creation, delta and completion of a Text content (a specific type of message). For convenience, the delta event includes both the incremental update and an accumulated snapshot of the content.
.on('imageFileDone', (content: ImageFile, snapshot: Message) => ...)
Image files are not sent incrementally so an event is provided for when a image file is available.
.on('toolCallCreated', (toolCall: ToolCall) => ...)
.on('toolCallDelta', (delta: RunStepDelta, snapshot: ToolCall) => ...)
.on('toolCallDone', (toolCall: ToolCall) => ...)
These events allow you to subscribe to events for the creation, delta and completion of a ToolCall.
More information on tools can be found here Tools
.on('end', () => ...)
The last event send when a stream ends.
The assistant streaming object also provides a few methods for convenience:
.currentEvent(): AssistantStreamEvent | undefined
.currentRun(): Run | undefined
.currentMessageSnapshot(): Message
.currentRunStepSnapshot(): Runs.RunStep
These methods are provided to allow you to access additional context from within event handlers. In many cases the handlers should include all the information you need for processing, but if additional context is required it can be accessed.
Note: There is not always a relevant context in certain situations (these will be undefined
in those cases).
await .finalMessages() : Promise<Message[]>
await .finalRunSteps(): Promise<RunStep[]>
These methods are provided for convenience to collect information at the end of a stream. Calling these events will trigger consumption of the stream until completion and then return the relevant accumulated objects.
openai.chat.completions.stream({ stream?: false, … }, options?): ChatCompletionStreamingRunner
openai.chat.completions.stream()
returns a ChatCompletionStreamingRunner
, which emits events, has an async
iterator, and exposes helper methods to accumulate chunks into a convenient shape and make it easy to reason
about the conversation.
Alternatively, you can use openai.chat.completions.create({ stream: true, … })
which returns an async
iterable of the chunks in the stream and uses less memory (most notably, it does not accumulate a final chat
completion object for you).
If you need to cancel a stream, you can break
from a for await
loop or call stream.abort()
.
See an example of streaming helpers in action in examples/stream.ts
.
openai.chat.completions.runTools({ stream: false, … }, options?): ChatCompletionRunner
openai.chat.completions.runTools({ stream: true, … }, options?): ChatCompletionStreamingRunner
openai.chat.completions.runTools()
returns a Runner
for automating function calls with chat completions.
The runner automatically calls the JavaScript functions you provide and sends their results back
to the API, looping as long as the model requests function calls.
If you pass a parse
function, it will automatically parse the arguments
for you and returns any parsing
errors to the model to attempt auto-recovery. Otherwise, the args will be passed to the function you provide
as a string.
client.chat.completions.runTools({
model: 'gpt-3.5-turbo',
messages: [{ role: 'user', content: 'How is the weather this week?' }],
tools: [
{
type: 'function',
function: {
function: getWeather as (args: { location: string; time: Date }) => any,
parse: parseFunction as (args: strings) => { location: string; time: Date },
parameters: {
type: 'object',
properties: {
location: { type: 'string' },
time: { type: 'string', format: 'date-time' },
},
},
},
},
],
});
If you pass function_call: {name: …}
instead of auto
, it returns immediately after calling that
function (and only loops to auto-recover parsing errors).
By default, we run the loop up to 10 chat completions from the API. You can change this behavior by
adjusting maxChatCompletions
in the request options object. Note that max_tokens
is the limit per
chat completion request, not for the entire call run.
See an example of automated function calls in action in
examples/function-call-helpers.ts
.
Note, runFunctions
was also previously available, but has been deprecated in favor of runTools
.
The first event that is fired when the connection with the OpenAI API is established.
The event fired when a chunk is received from the API. Not fired when it is not streaming. The snapshot
returns an accumulated ChatCompletionSnapshot
, which has a similar shape to ChatCompletion
with optional
fields and is built up from the chunks.
The event fired when a chat completion is returned or done being streamed by the API.
The event fired when a new message is either sent or received from the API. Does not fire for the messages
sent as the parameter to either .runTools()
or .stream()
The event fired when a message from the assistant
is received from the API.
The event fired when a chunk from the assistant
is received from the API. The delta
argument contains the
content of the chunk, while the snapshot
returns the accumulated content for the current message.
The event fired when a function call is made by the assistant.
The event fired when the function runner responds to the function call with role: "function"
. The content
of the
response is given as the first argument to the callback.
The event fired for every chunk containing new content. The props
object contains:
delta
: The new content string received in this chunksnapshot
: The accumulated content so farparsed
: The partially parsed content (if applicable)
The event fired when the content generation is complete. The props
object contains:
content
: The full generated contentparsed
: The fully parsed content (if applicable)
The event fired when a chunk contains part of a content refusal. The props
object contains:
delta
: The new refusal content string received in this chunksnapshot
: The accumulated refusal content string so far
The event fired when the refusal content is complete. The props
object contains:
refusal
: The full refusal content
The event fired when a chunk contains part of a function tool call's arguments. The props
object contains:
name
: The name of the function being calledindex
: The index of the tool callarguments
: The accumulated raw JSON string of argumentsparsed_arguments
: The partially parsed arguments objectarguments_delta
: The new JSON string fragment received in this chunk
The event fired when a function tool call's arguments are complete. The props
object contains:
name
: The name of the function being calledindex
: The index of the tool callarguments
: The full raw JSON string of argumentsparsed_arguments
: The fully parsed arguments object
The event fired when a chunk contains new content log probabilities. The props
object contains:
content
: A list of the new log probabilities received in this chunksnapshot
: A list of the accumulated log probabilities so far
The event fired when all content log probabilities have been received. The props
object contains:
content
: The full list of token log probabilities for the content
The event fired when a chunk contains new refusal log probabilities. The props
object contains:
refusal
: A list of the new log probabilities received in this chunksnapshot
: A list of the accumulated log probabilities so far
The event fired when all refusal log probabilities have been received. The props
object contains:
refusal
: The full list of token log probabilities for the refusal
The event fired for the final chat completion. If the function call runner exceeds the number
maxChatCompletions
, then the last chat completion is given.
The event fired for the content
of the last role: "assistant"
message. Not fired if there is no assistant
message.
The event fired for the last message.
The event fired for the last message with a defined function_call
.
The event fired for the last message with a role: "function"
.
The event fired when an error is encountered outside of a parse
function or an abort.
The event fired when the stream receives a signal to abort.
.on('totalUsage', (usage: CompletionUsage) => …)
(without stream
, usage is not currently reported with stream
)
The event fired at the end, returning the total usage of the call.
The last event fired in the stream.
Aborts the runner and the streaming request, equivalent to .controller.abort()
. Calling .abort()
on a
ChatCompletionStreamingRunner
will also abort any in-flight network requests.
An empty promise which resolves when the stream is done.
A promise which resolves with the final chat completion that was received from the API. Throws if the request ends before a complete chat completion is returned.
A promise which resolves with The array of all chat completions that were received from the API.
A promise which resolves with the content
of the last role: "assistant"
message. Throws if no such message
can be found.
A promise which resolves with the last message.
A promise which resolves with the last message with a defined function_call
. Throws if no such message is
found.
A promise which resolves with the last message with a role: "function"
. Throws if no such message is found.
A promise which resolves with the total usage.
A mutable array of all messages in the conversation.
The underlying AbortController
for the runner.
If you have a function call flow which you intend to end with a certain function call, then you can use the second
argument runner
given to the function to either mutate runner.messages
or call runner.abort()
.
import OpenAI from 'openai';
const client = new OpenAI();
async function main() {
const runner = client.chat.completions
.runTools({
model: 'gpt-3.5-turbo',
messages: [{ role: 'user', content: "How's the weather this week in Los Angeles?" }],
tools: [
{
type: 'function',
function: {
function: function updateDatabase(props, runner) {
runner.abort()
},
…
}
},
],
})
.on('message', (message) => console.log(message));
const finalFunctionCall = await runner.finalFunctionCall();
console.log('Final function call:', finalFunctionCall);
}
main();
zod
is a schema validation library which can help with validating the
assistant's response to make sure it conforms to a schema. Paired with zod-to-json-schema
, the validation schema also acts as the parameters
JSON Schema passed to the API.
import OpenAI from 'openai';
import { z } from 'zod';
import { zodToJsonSchema } from 'zod-to-json-schema';
const client = new OpenAI();
async function main() {
const runner = client.chat.completions
.runTools({
model: 'gpt-3.5-turbo',
messages: [{ role: 'user', content: "How's the weather this week in Los Angeles?" }],
tools: [
{
type: 'function',
function: {
function: getWeather,
parse: GetWeatherParameters.parse,
parameters: zodToJsonSchema(GetWeatherParameters),
},
},
],
})
.on('message', (message) => console.log(message));
const finalContent = await runner.finalContent();
console.log('Final content:', finalContent);
}
const GetWeatherParameters = z.object({
location: z.enum(['Boston', 'New York City', 'Los Angeles', 'San Francisco']),
});
async function getWeather(args: z.infer<typeof GetWeatherParameters>) {
const { location } = args;
// … do lookup …
return { temperature, precipitation };
}
main();
See a more fully-fledged example in examples/function-call-helpers-zod.ts
.
See an example of a Next.JS integration here examples/stream-to-client-next.ts
.
See an example of using express to stream to a browser here examples/stream-to-client-express.ts
.
When interacting with the API some actions such as starting a Run and adding files to vector stores are asynchronous and take time to complete.
The SDK includes helper functions which will poll the status until it reaches a terminal state and then return the resulting object.
If an API method results in an action which could benefit from polling there will be a corresponding version of the
method ending in _AndPoll
.
All methods also allow you to set the polling frequency, how often the API is checked for an update, via a function argument (pollIntervalMs
).
The polling methods are:
client.beta.threads.createAndRunPoll(...)
client.beta.threads.runs.createAndPoll((...)
client.beta.threads.runs.submitToolOutputsAndPoll((...)
client.beta.vectorStores.files.uploadAndPoll((...)
client.beta.vectorStores.files.createAndPoll((...)
client.beta.vectorStores.fileBatches.createAndPoll((...)
client.beta.vectorStores.fileBatches.uploadAndPoll((...)