DeepSeek

Configuration#

'deepseek' => [
    'api_key' => env('DEEPSEEK_API_KEY', ''),
    'url' => env('DEEPSEEK_URL', 'https://api.deepseek.com/v1')
]

Text, tools and reasoning#

use Prism\Prism\Facades\Prism;
use Prism\Prism\Facades\Tool;

$tool = Tool::as('multiply')
    ->for('Multiply two numbers')
    ->withNumberParameter('left', 'First factor')
    ->withNumberParameter('right', 'Second factor')
    ->withProviderOptions(['strict' => true])
    ->using(fn (float $left, float $right): string => (string) ($left * $right));

$response = Prism::text()
    ->using('deepseek', 'deepseek-flash')
    ->withTools([$tool])
    ->withMaxSteps(3)
    ->withPrompt('What is 12 times 13?')
    ->asText();

echo $response->text;
$reasoning = $response->additionalContent['reasoning_content'] ?? null;

Per-tool strict is forwarded in the function definition. Request provider options such as thinking and reasoning_effort are passed through to DeepSeek; explicit Prism model, messages, token and sampling settings take precedence over conflicting options. Use options supported by the selected model. See the DeepSeek model catalog for available IDs.

Structured output#

use Prism\Prism\Schema\ObjectSchema;
use Prism\Prism\Schema\StringSchema;

$response = Prism::structured()
    ->using('deepseek', 'deepseek-flash')
    ->withSchema(new ObjectSchema(
        'answer', 'An answer', [new StringSchema('text', 'Answer text')],
        requiredFields: ['text'],
    ))
    ->withPrompt('Describe a database index.')
    ->asStructured();

The structured handler uses json_object mode and adds schema instructions to a system message. It does not send a native JSON Schema constraint. Read decoded data from $response->structured and validate application-specific requirements.

Errors#

HTTP 402 raises PrismException with an Insufficient Balance message. Other failed HTTP requests include the provider's status and error details. See Error Handling.

Streaming#

DeepSeek supports streaming responses in real-time. All standard streaming methods are supported:

return Prism::text()
    ->using('deepseek', 'deepseek-flash')
    ->withPrompt(request('message'))
    ->asEventStreamResponse();

For complete streaming documentation, see Streaming Output.

Token Usage#

DeepSeek's prompt_tokens is the sum of its cache hits and cache misses, so Prism splits the hits out to keep the input buckets apart (see Token Usage). Text, structured and streamed responses all report usage like this:

When DeepSeek reports 0 cache hits, cacheReadInputTokens is 0 rather than null, so you can tell "no cache hit" apart from "not reported".

Limitations#

Embeddings#

Does not support embeddings.

Tool Choice#

Use withToolChoice() with ToolChoice::Auto or a tool name to control tool selection. This integration rejects ToolChoice::Any and ToolChoice::None.

Images#

The integration maps image parts into image URL or base64 content. Select a model that accepts image input; provider and model support determine whether the request succeeds.