Skip to content

Streaming Responses

Streaming responses let you process data as it arrives from the server rather than waiting for the entire response to buffer in memory. This is particularly valuable when working with LLM APIs that generate tokens incrementally, downloading large files, or consuming real-time event streams.

Streams are considered complete only after their iterator reaches the natural end. Breaking early leaves a stream incomplete; one-shot streams cannot then be replayed. When constructing a response from an iterable, use HttpResponse::streamingFromIterable() for non-buffering consumption or HttpResponse::bufferedFromIterable() when replay is explicitly required. Passing a raw iterable to HttpResponse is rejected so it cannot silently retain every chunk.

Enabling Streaming

To receive a streaming response, set the stream option on the request:

use Cognesy\Http\Data\HttpRequest;

$request = new HttpRequest(
    url: 'https://api.example.com/stream',
    method: 'GET',
    headers: ['Accept' => 'text/event-stream'],
    body: '',
    options: ['stream' => true],
);

You can also enable streaming on an existing request using withStreaming():

$request = $request->withStreaming(true);

Consuming the Stream

Once you have a streaming request, call stream() on the pending response. This returns a PHP Generator that yields string chunks:

foreach ($client->send($request)->stream() as $chunk) {
    echo $chunk;
    flush();
}

Each chunk is a raw string as received from the transport layer. The size of individual chunks depends on the driver and the streamChunkSize setting in HttpClientConfig (default: 256 bytes).

Note: You do not need to explicitly set stream => true on the request when using PendingHttpResponse::stream(). The pending response will force streaming mode automatically. However, setting it on the request is useful when middleware needs to know the intended mode before execution.

Streaming LLM Responses

Streaming is essential for AI/LLM integrations where responses are generated token by token. Here is a typical pattern for streaming a chat completion:

$request = new HttpRequest(
    url: 'https://api.openai.com/v1/chat/completions',
    method: 'POST',
    headers: [
        'Content-Type' => 'application/json',
        'Authorization' => 'Bearer ' . $apiKey,
    ],
    body: [
        'model' => 'gpt-4',
        'messages' => [
            ['role' => 'user', 'content' => 'Write a haiku about PHP.'],
        ],
        'stream' => true,
    ],
    options: ['stream' => true],
);

foreach ($client->send($request)->stream() as $chunk) {
    echo $chunk;
    flush();
}

The raw chunks from the transport layer will contain server-sent event framing (e.g., data: {...}\n\n). To parse these into clean payloads, use the EventSourceMiddleware.

Server-Sent Events with EventSourceMiddleware

The EventSourceMiddleware handles the SSE protocol for you. It strips the data: prefixes, buffers partial lines, and yields complete event payloads. A final event is emitted when the source ends even if it has no trailing blank line. The parser buffer is limited to 1 MiB by default; configure maxBufferBytes on EventSourceMiddleware for a smaller or larger protocol-specific limit:

use Cognesy\Http\Extras\Middleware\EventSource\EventSourceMiddleware;

$client = $client->withMiddleware(
    (new EventSourceMiddleware(true))
        ->withParser(fn(string $payload): string => $payload),
    'eventsource',
);

The parser callback receives the raw payload string from each data: line and returns the value to yield. Return false to skip an event. This is useful for filtering out [DONE] markers or parsing JSON:

$client = $client->withMiddleware(
    (new EventSourceMiddleware(true))
        ->withParser(function (string $payload): string|bool {
            if ($payload === '[DONE]') {
                return false; // skip
            }
            return $payload;
        }),
    'eventsource',
);

You can also attach listeners for debugging or event dispatching:

use Cognesy\Http\Extras\Support\EventSource\Listeners\PrintToConsole;
use Cognesy\Http\Config\DebugConfig;

$middleware = (new EventSourceMiddleware(true))
    ->withListeners(new PrintToConsole(new DebugConfig(httpEnabled: true)))
    ->withParser(fn(string $payload): string => $payload);

Downloading Large Files

Streaming is the right approach for downloading large files without exhausting memory:

$request = new HttpRequest(
    url: 'https://example.com/large-dataset.csv',
    method: 'GET',
    headers: [],
    body: '',
    options: ['stream' => true],
);

$handle = fopen('dataset.csv', 'wb');

foreach ($client->send($request)->stream() as $chunk) {
    fwrite($handle, $chunk);
}

fclose($handle);

Capturing stream contents for inspection

Streamed bodies are consumed once and normally leave no trace. When you need to inspect what actually came over the wire — debugging a malformed SSE stream, building replay tooling, capturing a postmortem sample — enable opt-in capture on the response before consuming it:

use Cognesy\Http\Stream\StreamCapturingPolicy;

$response = $client->send($request)
    ->get()
    ->withStreamCapture(StreamCapturingPolicy::preview()); // first 64KB

foreach ($response->stream() as $chunk) {
    // process chunks as usual — capture happens transparently
}

$capture = $response->streamCapture();
echo $capture->preview();          // captured prefix of the raw stream
$stats = $capture->stats();        // bytes, chunks, capturedBytes, truncated

Policies bound memory explicitly:

  • StreamCapturingPolicy::preview(int $maxBytes = 65536) — capture a prefix, enough to see what the stream looked like
  • StreamCapturingPolicy::chunks(int $maxBytes = 1048576) — retain individual chunks (via capturedChunks()) up to the byte budget
  • StreamCapturingPolicy::full(int $maxBytes) — capture everything up to an explicit cap (capturedBody() returns the concatenated content)
  • StreamCapturingPolicy::disabled() — pass-through, zero retention

Capture is per-response and off by default; it never changes what the consumer of stream() sees. The truncated flag in stats() tells you when the byte

The default is intentionally disabled because capture adds per-chunk work and retention to a path designed to stay one-shot. With 50,000–100,000 small chunks, chunks()/full() can retain 50,000–100,000 PHP string and array entries before their byte cap is reached; the object overhead can be several times larger than the payload itself. preview() retains only one bounded prefix, while disabled() avoids capture storage entirely. Choose a small preview for diagnostics on long-lived streams and enable full/chunk capture only when the retained data is explicitly needed.

Considerations

When working with streaming responses, keep these points in mind:

  • Memory usage. Streaming avoids buffering the entire response, but be careful not to accumulate chunks in a variable unless you actually need the full content.
  • Connection stability. Streaming connections stay open longer and are more sensitive to network interruptions. Pair streaming with retry middleware for resilience.
  • Timeouts. The idleTimeout setting in HttpClientConfig controls how long the client waits between data packets. Set it to -1 to disable idle timeouts for long-lived streams.
  • Body access. Calling body() on a streamed HttpResponse throws a LogicException. Always use stream() for streamed responses.
  • Middleware order. Middleware that decorates the stream (like EventSourceMiddleware) should be registered before middleware that reads the final content.