Skip to main content

​
Overview

Google offers Gemini models which perform well in benchmarks. Supported modes:
  • OutputMode::MdJson - fallback mode
  • OutputMode::Json - recommended
  • OutputMode::Tools - supported
Here’s how you can use Instructor with Gemini API in OpenAI-compatible mode. 
<?php

use Cognesy\Http\Creation\HttpClientBuilder;
use Cognesy\Polyglot\Inference\Inference;
use Cognesy\Polyglot\Inference\InferenceRuntime;
use Cognesy\Utils\Str;
use Cognesy\Polyglot\Inference\Config\LLMConfig;

require 'examples/boot.php';

$http = (new HttpClientBuilder())->withDebugConfig(ExampleConfig::debugPreset('detailed'))->create();

$answer = Inference::fromRuntime(InferenceRuntime::fromConfig(
        config: ExampleConfig::llmPreset('gemini-oai'), // use OpenAI-compatible Gemini config (v1beta/openai)
        httpClient: $http,
    ))
    ->with(
        messages: [['role' => 'user', 'content' => 'What is the capital of France']],
        options: ['max_tokens' => 64]
    )
    ->get();

echo "USER: What is capital of France\n";
echo "ASSISTANT: $answer\n";
assert(Str::contains($answer, 'Paris'));
?>