<?php
// 安装依赖：composer require smalot/pdfbox
use Smalot\Pdf\Parser;class KnowledgeProcessor {private $embeddingUrl = "http://localhost:5000/embed";public function processDocument(string $filePath): array {$text = $this->extractText($filePath);$chunks = $this->chunkText($text);return $this->generateVectors($chunks);}private function extractText(string $path): string {$ext = pathinfo($path, PATHINFO_EXTENSION);switch (strtolower($ext)) {case 'pdf':$parser = new Parser();$pdf = $parser->parseFile($path);return $pdf->getText();case 'docx':return $this->readDocx($path);// 其他格式处理...}}private function chunkText(string $text, int $chunkSize=512): array {// 按语义分块（简化版）$sentences = preg_split('/(?<=[。！？])/u', $text);$chunks = [];$current = '';foreach ($sentences as $sentence) {if (mb_strlen($current) + mb_strlen($sentence) > $chunkSize) {$chunks[] = $current;$current = '';}$current .= $sentence;}return $chunks;}private function generateVectors(array $chunks): array {$vectors = [];foreach ($chunks as $index => $text) {$data = ['text' => $text];$options = ['http' => ['method' => 'POST','header' => 'Content-Type: application/json','content' => json_encode($data)]];$context = stream_context_create($options);$response = file_get_contents($this->embeddingUrl, false, $context);$vectors[] = ['text' => $text,'vector' => json_decode($response, true)['vector'],'chunk_id' => uniqid('chunk_')];}return $vectors;}
}// 使用示例
$processor = new KnowledgeProcessor();
$vectors = $processor->processDocument('/path/to/企业章程.pdf');

<?php
class KnowledgeBot {private $db;private $llmEndpoint = "http://localhost:8000/v1/chat/completions";public function __construct(PDO $db) {$this->db = $db;}public function query(string $question): string {// 1. 检索相关文本块$contexts = $this->retrieveContext($question);// 2. 构建提示词$prompt = $this->buildPrompt($question, $contexts);// 3. 调用大模型return $this->callLLM($prompt);}private function retrieveContext(string $query): array {// 获取查询向量（实际需调用Embedding服务）$queryVector = $this->getQueryVector($query);// 向量相似度搜索$stmt = $this->db->prepare("SELECT text, chunk_id FROM knowledge_vectors ORDER BY vector <=> :vector LIMIT 3");$stmt->execute([':vector' => json_encode($queryVector)]);return $stmt->fetchAll(PDO::FETCH_ASSOC);}private function buildPrompt(string $question, array $contexts): string {$contextText = "";foreach ($contexts as $ctx) {$contextText .= "【相关文档{$ctx['chunk_id']}】\n{$ctx['text']}\n\n";}return <<<PROMPT
你是一家企业的智能知识库助手，请根据以下上下文回答问题：
{$contextText}
---
用户问题：{$question}
回答要求：
1. 如果问题与提供资料无关，回答"该问题不在知识库范围内"
2. 引用文档时标注来源ID
3. 用中文简洁回答
PROMPT;}private function callLLM(string $prompt): string {$data = ['model' => 'chatglm3','messages' => [['role' => 'user', 'content' => $prompt]]];$options = ['http' => ['method' => 'POST','header' => "Content-Type: application/json",'content' => json_encode($data)]];$context = stream_context_create($options);$response = json_decode(file_get_contents($this->llmEndpoint, false, $context), true);return $response['choices'][0]['message']['content'];}
}// 使用示例
$db = new PDO('sqlite:knowledge.db');
$bot = new KnowledgeBot($db);
echo $bot->query("公司年假政策是怎样的？");

class TechnicalTokenizer {private $dict = [];public function __construct(array $terms) {// 加载专业词典$this->dict = array_flip($terms);}public function tokenize(string $text): array {$tokens = [];$length = mb_strlen($text);$pos = 0;while ($pos < $length) {$found = false;// 优先匹配长术语（最大正向匹配）for ($len = 10; $len >= 1; $len--) {if ($pos + $len > $length) continue;$chunk = mb_substr($text, $pos, $len);if (isset($this->dict[$chunk])) {$tokens[] = $chunk;$pos += $len;$found = true;break;}}if (!$found) {// 默认单字切分$tokens[] = mb_substr($text, $pos, 1);$pos++;}}return $tokens;}
}// 使用示例
$tokenizer = new TechnicalTokenizer(["晶圆蚀刻", "光刻胶", "FinFET"]);
print_r($tokenizer->tokenize("晶圆蚀刻工艺中光刻胶的使用规范"));
// 输出: ["晶圆蚀刻", "工", "艺", "中", "光刻胶", "的", "使", "用", "规", "范"]

class FactChecker {public function verify(string $answer, array $contexts): bool {// 1. 关键词匹配验证$keywords = $this->extractKeywords($answer);$matchScore = 0;foreach ($contexts as $ctx) {foreach ($keywords as $kw) {if (mb_strpos($ctx['text'], $kw) !== false) $matchScore++;}}if ($matchScore < count($keywords) * 0.6) return false;// 2. 语义相似度验证$answerVector = $this->getEmbedding($answer);$maxSim = 0;foreach ($contexts as $ctx) {$sim = $this->cosineSimilarity($answerVector, $ctx['vector']);if ($sim > $maxSim) $maxSim = $sim;}if ($maxSim < 0.75) return false;// 3. 逻辑自洽验证return $this->logicCheck($answer);}private function logicCheck(string $text): bool {// 使用规则引擎检查矛盾语句$contradictions = ['/虽然.*但是.*然而/' => 0.8,'/一方面.*另一方面/' => 0.6];$score = 1.0;foreach ($contradictions as $pattern => $penalty) {if (preg_match($pattern, $text)) {$score *= $penalty;}}return $score > 0.7;}
}

// 模型量化加载（使用llama.cpp）
$model = new LlamaCPP(['model_path' => '/models/qwen1.5-14b.Q4_K_M.gguf','n_gpu_layers' => 35,  // GPU加速层数'n_threads' => 8       // CPU线程数
]);// 流式输出实现
header('Content-Type: text/event-stream');
header('Cache-Control: no-cache');$prompt = "请解释量子计算原理";
$max_tokens = 300;
$chunk_size = 50;for ($i = 0; $i < $max_tokens; $i += $chunk_size) {$response = $model->generate($prompt, ['max_tokens' => $chunk_size,'stream' => true]);echo "data: " . json_encode(['text' => $response]) . "\n\n";ob_flush();flush();$prompt .= $response;  // 增量生成
}

# Nginx配置示例
location /api/ {proxy_pass http://backend;# 安全防护limit_req zone=apilimit burst=20;add_header Strict-Transport-Security "max-age=31536000";proxy_set_header X-Real-IP $remote_addr;# 文档类型白名单if ($request_filename ~* ^.*?\.(php|exe|sh)$) {return 403;}
}

// Milvus向量检索示例
$client = new Milvus\Grpc\MilvusServiceClient('localhost:19530');
$response = $client->search(['collection_name' => 'enterprise_knowledge','vectors' => [$queryVector],'top_k' => 5
]);