如何重构PHP代码使GPT-4生成连贯完整的5000字文档

原代码核心问题 #

1. 续写请求未传递有效指令：后续请求中用户消息为空，模型无法判断需要延续之前的内容，导致生成主题变体
2. 参数使用错误：Chat Completions接口不支持独立的prompt参数，需通过messages数组维护上下文逻辑
3. 长度控制不准确：用strlen统计字符数代替单词数，直接截断易破坏语句完整性

重构后的代码 #

public function writePrompt(Request $request)
{
    $this->authCheck();
    $this->validate($request, [
        'content' => 'required',
        'maxWords' => 'required|integer|min:1',
    ]);

    ini_set('max_execution_time', 1800);

    $userPrompt = $request->input('content');
    $targetWordCount = intval($request->input('maxWords'));
    $maxTokensPerRequest = 1500; // 预留token给指令，避免内容截断
    $generatedContent = '';
    $messageHistory = [
        [
            'role' => 'user',
            'content' => $userPrompt . "\n请生成一篇连贯完整的长文档，严格围绕主题展开，后续续写时保持内容的连续性和逻辑性。"
        ]
    ];

    if (empty($this->super_settings['openai_api_key']) || $this->isDemo() || $this->super_settings['openai_api_key'] == 'demo') {
        $generatedContent = __('Sorry, I am not able to write anything for you.');
        return $this->returnJsonResponse($generatedContent);
    }

    $client = OpenAI::client($this->super_settings['openai_api_key']);

    try {
        do {
            $response = $client->chat()->create([
                'model' => 'gpt-4',
                'messages' => $messageHistory,
                'max_tokens' => $maxTokensPerRequest,
                'temperature' => 0.7, // 平衡创造性与内容连贯性
            ]);

            $newContent = $response->choices[0]->message->content;
            $generatedContent .= $newContent;

            // 更新对话历史，让模型明确已有内容，后续继续续写
            $messageHistory[] = ['role' => 'assistant', 'content' => $newContent];
            $messageHistory[] = [
                'role' => 'user',
                'content' => "继续围绕之前的主题和内容续写，保持连贯性，直到总字数达到约{$targetWordCount}字。当前已生成的最近内容：" . substr($generatedContent, -500)
            ];

            $currentWordCount = $this->countWords($generatedContent);
        } while ($currentWordCount < $targetWordCount && $currentWordCount + 500 < $targetWordCount);

        // 调整内容到目标字数，优先保证语句完整
        $generatedContent = $this->trimToWordCount($generatedContent, $targetWordCount);
    } catch (\Exception $e) {
        $generatedContent = $this->user->is_super_admin ? 'Error: ' . $e->getMessage() : __('Sorry, I am not able to write anything for you.');
    }

    $generatedContent = Str::markdown($generatedContent);
    return $this->returnJsonResponse($generatedContent);
}

/**
 * 统计内容单词数（兼容中英文）
 */
private function countWords(string $content): int
{
    $chineseChars = preg_match_all('/[\x{4e00}-\x{9fa5}]/u', $content, $matches);
    $englishWords = str_word_count(preg_replace('/[\x{4e00}-\x{9fa5}]/u', '', $content));
    return $chineseChars + $englishWords;
}

/**
 * 将内容截断到指定单词数，尽量保证语句完整
 */
private function trimToWordCount(string $content, int $targetCount): string
{
    $currentCount = $this->countWords($content);
    if ($currentCount <= $targetCount) {
        return $content;
    }

    $punctuations = ['。', '！', '？', '.', '!', '?', ';', '；'];
    $trimmed = $content;
    while ($this->countWords($trimmed) > $targetCount) {
        $lastPos = PHP_INT_MAX;
        foreach ($punctuations as $punc) {
            $pos = strrpos($trimmed, $punc);
            if ($pos !== false && $pos < $lastPos) {
                $lastPos = $pos;
            }
        }
        if ($lastPos === PHP_INT_MAX) {
            $words = preg_split('/\s+|(?<=[\x{4e00}-\x{9fa5}])|(?=[\x{4e00}-\x{9fa5}])/u', $trimmed);
            $trimmed = implode('', array_slice($words, 0, $targetCount));
            break;
        }
        $trimmed = substr($trimmed, 0, $lastPos + 1);
    }
    return $trimmed;
}

/**
 * 返回统一JSON响应
 */
private function returnJsonResponse(string $content): \Illuminate\Http\JsonResponse
{
    return response()->json([
        'success' => true,
        'result' => $content,
    ]);
}

关键改进说明 #

• 上下文逻辑修复：通过messageHistory数组完整保存对话流程，每次续写时明确告知模型基于已有内容延续
• 指令明确化：初始请求和续写请求都添加清晰要求，强制模型保持内容连贯性
• 字数控制优化：新增中英文兼容的单词统计方法，截断时优先按标点分割，避免破坏语句结构
• 代码结构优化：拆分工具方法，提升代码可读性和维护性

内容的提问来源于stack exchange，提问作者Kevin Otieno