[
'bold_single' => 50,
'bold_multi' => 10,
'title.text.pali' => 3,
'title.text.zh' => 3,
'summary.text' => 2,
'content.text.pali' => 1,
'content.text.zh' => 1,
],
'hybrid' => [
'fuzzy_ratio' => 0.7,
'semantic_ratio' => 0.3,
'bold_single' => 50,
'bold_multi' => 10,
'title.text.pali' => 3,
'title.text.zh' => 3,
'summary.text' => 2,
'content.text.pali' => 1,
'content.text.zh' => 1,
],
];
/**
* OpenSearch 索引定义(settings + mappings)
*
* 字段结构说明:
*
* title
* ├── text
* │ ├── pali (text) 模糊查询 + exact subfield 精确查询
* │ └── zh (text) 中文分词查询
* ├── vector (knn_vector, dim=1536)
* └── suggest
* ├── pali (completion)
* └── zh (completion)
*
* content(结构与 title 一致,额外包含 tokens nested 字段)
* ├── text
* │ ├── pali (text)
* │ └── zh (text)
* ├── tokens (nested)
* ├── vector (knn_vector, dim=1536)
* └── suggest
* ├── pali (completion)
* └── zh (completion)
*
* summary(中文摘要,结构保持不变)
* ├── text (text)
* └── vector (knn_vector, dim=1536)
*
* @var array
*/
private $indexDefinition = [
'settings' => [
'index' => [
'knn' => true,
],
'analysis' => [
'analyzer' => [
'pali_query_analyzer' => [
'tokenizer' => 'standard',
'filter' => ['lowercase', 'pali_synonyms'],
],
'pali_index_analyzer' => [
'type' => 'custom',
'tokenizer' => 'standard',
'char_filter' => ['markdown_strip'],
'filter' => ['lowercase'],
],
'markdown_clean' => [
'type' => 'custom',
'tokenizer' => 'standard',
'char_filter' => ['markdown_strip'],
'filter' => ['lowercase'],
],
// Suggest 专用(忽略大小写 + 变音)
'pali_suggest_analyzer' => [
'tokenizer' => 'standard',
'filter' => ['lowercase', 'asciifolding'],
],
'zh_suggest_analyzer' => [
'tokenizer' => 'ik_max_word',
'char_filter' => ['tsconvert'],
],
// 中文简繁统一 (繁 -> 简)
'zh_index_analyzer' => [
'tokenizer' => 'ik_max_word',
'char_filter' => ['tsconvert'],
],
'zh_query_analyzer' => [
'tokenizer' => 'ik_smart',
'char_filter' => ['tsconvert'],
],
],
'filter' => [
'pali_synonyms' => [
'type' => 'synonym_graph',
'synonyms_path' => 'analysis/pali_synonyms.txt',
'updateable' => true,
],
],
'char_filter' => [
'markdown_strip' => [
'type' => 'pattern_replace',
'pattern' => '\\*\\*|\\*|_|`|~',
'replacement' => '',
],
'tsconvert' => [
'type' => 'stconvert',
'convert_type' => 't2s',
],
],
],
],
'mappings' => [
'properties' => [
'id' => ['type' => 'keyword'],
// 分类维度一律 keyword:它们是枚举值,不该被分词。
// 若声明成 text,'zh-Hans' 会被切成 zh + hans,查 language=zh 就会把
// zh-Hans 的文档一并捞进来(实测 401 vs 410),而且不报错。
'resource_id' => ['type' => 'keyword'],
'resource_type' => ['type' => 'keyword'],
// ----------------------------------------------------------------
// title
// text.pali → 模糊查询(+ exact subfield 精确查询)
// text.zh → 中文查询
// vector → 语义向量
// suggest.pali / suggest.zh → 自动建议
// ----------------------------------------------------------------
'title' => [
'properties' => [
'text' => [
'properties' => [
'pali' => [
'type' => 'text',
'analyzer' => 'pali_index_analyzer',
'search_analyzer' => 'pali_query_analyzer',
'fields' => [
'exact' => [
'type' => 'text',
'analyzer' => 'markdown_clean',
],
],
],
'zh' => [
'type' => 'text',
'analyzer' => 'zh_index_analyzer',
'search_analyzer' => 'zh_query_analyzer',
],
],
],
'vector' => [
'type' => 'knn_vector',
'dimension' => 1536,
'method' => [
'name' => 'hnsw',
'space_type' => 'innerproduct',
'engine' => 'faiss',
],
],
'suggest' => [
'properties' => [
'pali' => [
'type' => 'completion',
'analyzer' => 'pali_suggest_analyzer',
],
'zh' => [
'type' => 'completion',
'analyzer' => 'zh_suggest_analyzer',
],
],
],
],
],
// ----------------------------------------------------------------
// summary(LLM 生成的简体中文摘要,结构保持不变)
// text → 中文查询
// vector → 语义向量
// ----------------------------------------------------------------
'summary' => [
'properties' => [
'text' => [
'type' => 'text',
'analyzer' => 'zh_index_analyzer',
'search_analyzer' => 'zh_query_analyzer',
],
'vector' => [
'type' => 'knn_vector',
'dimension' => 1536,
'method' => [
'name' => 'hnsw',
'space_type' => 'innerproduct',
'engine' => 'faiss',
],
],
],
],
// ----------------------------------------------------------------
// content(结构与 title 对称,额外包含 tokens nested 字段)
// text.pali → 模糊查询(+ exact subfield 精确查询)
// text.zh → 中文查询
// tokens → 词法分析结果(nested)
// vector → 语义向量
// suggest.pali / suggest.zh → 自动建议
// ----------------------------------------------------------------
'content' => [
'properties' => [
'text' => [
'properties' => [
'pali' => [
'type' => 'text',
'analyzer' => 'pali_index_analyzer',
'search_analyzer' => 'pali_query_analyzer',
'fields' => [
'exact' => [
'type' => 'text',
'analyzer' => 'markdown_clean',
],
],
],
'zh' => [
'type' => 'text',
'analyzer' => 'zh_index_analyzer',
'search_analyzer' => 'zh_query_analyzer',
],
],
],
'tokens' => [
'type' => 'nested',
'properties' => [
'surface' => ['type' => 'keyword'],
'lemma' => ['type' => 'keyword'],
'compound_parts' => ['type' => 'keyword'],
'case' => ['type' => 'keyword'],
],
],
'vector' => [
'type' => 'knn_vector',
'dimension' => 1536,
'method' => [
'name' => 'hnsw',
'space_type' => 'innerproduct',
'engine' => 'faiss',
],
],
'suggest' => [
'properties' => [
'pali' => [
'type' => 'completion',
'analyzer' => 'pali_suggest_analyzer',
],
'zh' => [
'type' => 'completion',
'analyzer' => 'zh_suggest_analyzer',
],
],
],
// 前端展示用,原始 HTML,不参与索引
'display' => [
'type' => 'text',
'index' => false,
],
],
],
'related_id' => ['type' => 'keyword'],
'bold_single' => [
'type' => 'text',
'analyzer' => 'standard',
'search_analyzer' => 'pali_query_analyzer',
],
'bold_multi' => [
'type' => 'text',
'analyzer' => 'standard',
'search_analyzer' => 'pali_query_analyzer',
],
'path' => ['type' => 'text', 'analyzer' => 'standard'],
'page_refs' => ['type' => 'keyword'],
'tags' => ['type' => 'keyword'],
'category' => ['type' => 'keyword'],
'author' => ['type' => 'text'],
'language' => ['type' => 'keyword'],
'updated_at' => ['type' => 'date'],
'granularity' => ['type' => 'keyword'],
'metadata' => [
'properties' => [
'APA' => ['type' => 'text', 'index' => false],
'MLA' => ['type' => 'text', 'index' => false],
'widget' => ['type' => 'text', 'index' => false],
'author' => ['type' => 'text'],
'channel' => ['type' => 'text'],
],
],
],
],
];
/**
* 创建 OpenSearchService 实例
*
* 从 config('mint.opensearch.config') 读取连接配置,
* 同时初始化 OpenAI HTTP 客户端用于 embedding 调用。
*/
public function __construct()
{
$config = config('mint.opensearch.config');
$hostUrl = "{$config['scheme']}://{$config['host']}:{$config['port']}";
$this->client = (new GuzzleClientFactory)->create([
'base_uri' => $hostUrl,
'auth' => [$config['username'], $config['password']],
'verify' => $config['ssl_verification'],
]);
$this->openaiApiKey = env('OPENAI_API_KEY');
$this->http = new Client([
'base_uri' => 'https://api.openai.com/v1/',
'timeout' => 15,
]);
}
/**
* 动态覆盖指定搜索模式的字段权重
*
* @param string $mode 搜索模式,支持 'fuzzy' | 'hybrid'
* @param array $weights 需要覆盖的权重键值对,例如:['title.text.pali' => 5]
*/
public function setWeights(string $mode, array $weights): void
{
if (isset($this->weights[$mode])) {
$this->weights[$mode] = array_merge($this->weights[$mode], $weights);
}
}
/**
* 测试与 OpenSearch 集群的连接状态
*
* @return array{0: bool, 1: string} [连接是否成功, 描述信息]
*/
public function testConnection(): array
{
try {
$info = $this->client->info();
$message = 'OpenSearch 连接成功: '.json_encode($info['version']['number']);
Log::info($message);
return [true, $message];
} catch (Exception $e) {
$message = 'OpenSearch 连接失败: '.$e->getMessage();
Log::error($message);
return [false, $message];
}
}
/**
* 检查当前索引是否已存在
*/
public function indexExists(): bool
{
$index = config('mint.opensearch.index');
return $this->client->indices()->exists(['index' => $index]);
}
/**
* 创建 OpenSearch 索引
*
* 使用 $indexDefinition 中定义的 settings 和 mappings 创建索引。
* 若索引已存在则抛出异常,避免覆盖生产数据。
*
* @return array OpenSearch 响应
*
* @throws Exception 索引已存在时抛出
*/
public function createIndex(): array
{
$index = config('mint.opensearch.index');
$exists = $this->client->indices()->exists(['index' => $index]);
if ($exists) {
throw new Exception("Index [$index] already exists.");
}
return $this->client->indices()->create([
'index' => $index,
'body' => $this->indexDefinition,
]);
}
/**
* 更新已有索引的 settings 和 mappings
*
* 更新 settings 时会临时关闭索引(close → putSettings → open),
* 更新 mappings 支持热更新(新增字段),不可修改已有字段类型。
*
* @return array 包含 'settings' 和/或 'mappings' 的响应数组
*/
public function updateIndex(): array
{
$index = config('mint.opensearch.index');
$settings = $this->indexDefinition['settings'] ?? [];
$mappings = $this->indexDefinition['mappings'] ?? [];
$response = [];
if (! empty($settings)) {
$this->client->indices()->close(['index' => $index]);
$response['settings'] = $this->client->indices()->putSettings([
'index' => $index,
'body' => ['settings' => $settings],
]);
$this->client->indices()->open(['index' => $index]);
}
if (! empty($mappings)) {
$response['mappings'] = $this->client->indices()->putMapping([
'index' => $index,
'body' => $mappings,
]);
}
return $response;
}
/**
* 切换 pali 同义词文件版本
*
* 把 pali_synonyms filter 的 synonyms_path 指向
* analysis/pali-synonyms-{$version}.txt(文件须已存在于 OpenSearch
* 各节点的 config 目录中,否则索引无法重新打开)。
*
* synonyms_path 属于静态 settings,必须 close → putSettings → open。
* 无论 putSettings 成功与否都会尝试重新打开索引,避免索引停留在 close 状态。
*
* @param string $version 版本号,仅允许 [A-Za-z0-9._-]
* @return array{path: string, settings: array} 新路径与 OpenSearch 响应
*
* @throws Exception 版本号非法、索引不存在或 OpenSearch 拒绝时抛出
*
* @example
* $service->updatePaliSynonymsPath('20260731');
*/
public function updatePaliSynonymsPath(string $filename): array
{
$index = config('mint.opensearch.index');
if (! $this->client->indices()->exists(['index' => $index])) {
throw new Exception("Index [$index] does not exist.");
}
$path = Str::finish(config('mint.opensearch.config.synonyms_path'), '/').$filename;
Log::debug('OpenSearchService::updatePaliSynonymsPath', [
'index' => $index,
'synonyms_path' => $path,
]);
// 以代码中的 analysis 定义为准,只替换同义词文件路径
$analysis = $this->indexDefinition['settings']['analysis'];
$analysis['filter']['pali_synonyms']['synonyms_path'] = $path;
$this->client->indices()->close(['index' => $index]);
try {
$response = $this->client->indices()->putSettings([
'index' => $index,
'body' => ['settings' => ['analysis' => $analysis]],
]);
} finally {
$this->client->indices()->open(['index' => $index]);
}
return ['path' => $path, 'settings' => $response];
}
/**
* 读取当前索引使用的 pali 同义词文件路径
*
* @return string|null 形如 "analysis/pali-synonyms-20260731.txt",未设置时返回 null
*/
public function getPaliSynonymsPath(): ?string
{
return $this->getPaliSynonymsSetting()['synonyms_path'] ?? null;
}
/**
* 读取当前索引 pali_synonyms filter 的完整设置
*
* @return array{type?: string, synonyms_path?: string, updateable?: string}|null 未设置时返回 null
*/
public function getPaliSynonymsSetting(): ?array
{
$index = config('mint.opensearch.index');
$settings = $this->client->indices()->getSettings(['index' => $index]);
return $settings[$index]['settings']['index']['analysis']['filter']['pali_synonyms'] ?? null;
}
/**
* 删除当前索引
*
* @return array OpenSearch 响应
*/
public function deleteIndex(): array
{
$index = config('mint.opensearch.index');
return $this->client->indices()->delete(['index' => $index]);
}
/**
* 统计索引文档数量(支持可选条件过滤)
*
* @param array|null $query OpenSearch DSL query 子句,为 null 时统计全部文档。
* 示例:['term' => ['language' => 'zh']]
* ['exists' => ['field' => 'content.vector']]
* @return int 文档总数
*
* @throws Exception
*
* @example
* $service->count();
* $service->count(['exists' => ['field' => 'content.vector']]);
*/
public function count(?array $query = null): int
{
$index = config('mint.opensearch.index');
$params = ['index' => $index];
if (! empty($query)) {
$params['body'] = ['query' => $query];
}
$response = $this->client->count($params);
return (int) ($response['count'] ?? 0);
}
/**
* 写入或覆盖单条文档
*
* @param string $id 文档 ID
* @param array $body 文档内容,字段结构须与 mappings 一致
* @return array OpenSearch 响应
*/
public function create(string $id, array $body): array
{
return $this->client->index([
'index' => config('mint.opensearch.index'),
'id' => $id,
'body' => $body,
]);
}
/**
* 删除单条文档
*
* @param string $id 文档 ID
* @return array OpenSearch 响应
*/
public function delete(string $id): array
{
return $this->client->delete([
'index' => config('mint.opensearch.index'),
'id' => $id,
]);
}
/**
* 执行高级搜索
*
* 支持五种搜索模式:
* - fuzzy 多字段模糊查询(默认),巴利文走三层排序
* - exact 精确匹配,使用 markdown_clean analyzer
* - semantic 纯语义向量搜索,需要 OpenAI embedding
* - hybrid fuzzy + semantic 混合,权重由 fuzzy_ratio / semantic_ratio 控制
* - definition 定义查询,黑体字(bold_single/bold_multi)高权重,术语定义排前
*
* 支持的过滤参数:
* resourceType, resourceId, granularity, language, category,
* tags, pageRefs, relatedId, author, channel
*
* @param array $params {
*
* @type string $query 搜索关键词(必填)
* @type string $searchMode 搜索模式,默认 'fuzzy'
* @type int $page 页码,默认 1
* @type int $pageSize 每页条数,默认 20
* @type string $resourceType 按资源类型过滤
* @type string $resourceId 按资源 ID 过滤
* @type string $granularity 按粒度过滤
* @type string $language 按语言过滤
* @type string $category 按分类过滤
* @type array $tags 按标签过滤(terms)
* @type array $pageRefs 按页码引用过滤(terms)
* @type string $relatedId 按关联 ID 过滤
* @type string $author 按作者过滤
* @type string $channel 按频道过滤
* @type array $highlight_pre_tags 高亮前置标签,默认 ['']
* @type array $highlight_post_tags 高亮后置标签,默认 ['']
* }
*
* @return array OpenSearch 原始响应
*
* @throws Exception semantic / hybrid 模式下 embedding 调用失败时抛出
*/
public function search(array $params): array
{
$page = $params['page'] ?? 1;
$pageSize = $params['pageSize'] ?? 20;
$from = ($page - 1) * $pageSize;
$mode = $params['searchMode'] ?? 'fuzzy';
// 排除字段
if (! empty($params['excludes']) && is_array($params['excludes'])) {
$excludes = array_merge($this->sourceExcludes, $params['excludes']);
} else {
$excludes = $this->sourceExcludes;
}
// ---------- 过滤条件 ----------
$filters = [];
if (! empty($params['resourceType'])) {
$filters[] = ['term' => ['resource_type' => $params['resourceType']]];
}
if (! empty($params['resourceId'])) {
$filters[] = ['term' => ['resource_id' => $params['resourceId']]];
}
if (! empty($params['granularity'])) {
$filters[] = ['term' => ['granularity' => $params['granularity']]];
}
if (! empty($params['language'])) {
$filters[] = ['term' => ['language' => $params['language']]];
}
if (! empty($params['category'])) {
if (is_array($params['category'])) {
$categories = $params['category'];
} else {
$categories = [$params['category']];
}
// 必须匹配全部:为每个 category 创建一个 term 条件
foreach ($categories as $category) {
$filters[] = ['term' => ['category' => $category]];
}
}
if (! empty($params['tags'])) {
$filters[] = ['terms' => ['tags' => $params['tags']]];
}
if (! empty($params['pageRefs'])) {
$filters[] = ['terms' => ['page_refs' => $params['pageRefs']]];
}
if (! empty($params['relatedId'])) {
$filters[] = ['term' => ['related_id' => $params['relatedId']]];
}
if (! empty($params['author'])) {
$filters[] = ['match' => ['metadata.author' => $params['author']]];
}
if (! empty($params['channel'])) {
$filters[] = ['term' => ['metadata.channel' => $params['channel']]];
}
// ---------- 查询部分 ----------
$queryText = trim($params['query'] ?? '');
// fuzzy / definition 的巴利文多词查询按词的罕见度加权;其余模式不需要
$weights = null;
if ($queryText !== '' && ! in_array($mode, ['exact', 'semantic', 'hybrid'], true) && ! $this->isCjkQuery($queryText)) {
$weights = $this->wordRarityWeights($this->scoringWords($this->splitQueryWords($queryText)));
}
if ($queryText === '') {
$query = ['match_all' => new \stdClass];
} else {
switch ($mode) {
case 'exact':
$query = $this->buildExactQuery($queryText);
break;
case 'semantic':
$query = $this->buildSemanticQuery($queryText);
break;
case 'hybrid':
$query = $this->buildHybridQuery($queryText);
break;
case 'definition':
$query = $this->buildDefinitionQuery($queryText, $weights);
break;
case 'fuzzy':
default:
$query = $this->buildFuzzyQuery($queryText, $weights);
break;
}
}
$highlightPreTags = $params['highlight_pre_tags'] ?? [''];
$highlightPostTags = $params['highlight_post_tags'] ?? [''];
// ---------- 最终 DSL ----------
$dsl = [
'from' => $from,
'size' => $pageSize,
'_source' => ['excludes' => $excludes],
'query' => ! empty($filters)
? [
'bool' => [
'must' => [$query],
'filter' => $filters,
],
]
: $query,
'aggs' => [
'resource_type' => [
'terms' => ['field' => 'resource_type'],
],
'language' => [
'terms' => ['field' => 'language'],
],
'category' => [
'terms' => ['field' => 'category'],
],
'granularity' => [
'terms' => ['field' => 'granularity'],
],
],
];
// 只有有搜索词时才开启高亮;字段按语种路由,避免把巴利文喂给中文分词器
if ($queryText !== '') {
$highlightFields = $this->isCjkQuery($queryText)
? ['title.text.zh', 'summary.text', 'content.text.zh']
: ['title.text.pali', 'content.text.pali'];
$dsl['highlight'] = [
'fields' => array_fill_keys($highlightFields, new \stdClass),
'fragmenter' => 'sentence',
'fragment_size' => 200,
'number_of_fragments' => 1,
'pre_tags' => $highlightPreTags,
'post_tags' => $highlightPostTags,
];
}
Log::debug(
'OpenSearchService::search',
['dsl' => json_encode($dsl, JSON_UNESCAPED_UNICODE)]
);
return $this->client->search([
'index' => config('mint.opensearch.index'),
'body' => $dsl,
]);
}
/**
* 构建 exact(精确匹配)查询
*
* 使用 markdown_clean analyzer 的 exact subfield 进行匹配,
* 适合巴利文词形精确检索场景。
*
* 查询字段:title.text.pali.exact, content.text.pali.exact, summary.text
*
* @param string $query 搜索关键词
* @return array OpenSearch DSL query 片段
*/
protected function buildExactQuery(string $query): array
{
return [
'multi_match' => [
'query' => $query,
'fields' => [
'title.text.pali.exact',
'content.text.pali.exact',
'summary.text',
],
'type' => 'best_fields',
],
];
}
/**
* 构建 semantic(纯语义向量)查询
*
* 将查询文本通过 OpenAI embedding API 转为向量,
* 同时对 content.vector、summary.vector、title.vector 三个 knn 字段检索,
* 使用 bool should 合并结果。
*
* @param string $query 搜索关键词
* @return array OpenSearch DSL query 片段
*
* @throws Exception embedding 调用失败时抛出
*/
protected function buildSemanticQuery(string $query): array
{
$vector = $this->embedText($query);
return [
'bool' => [
'should' => [
['knn' => ['content.vector' => ['vector' => $vector, 'k' => 20]]],
['knn' => ['summary.vector' => ['vector' => $vector, 'k' => 10]]],
['knn' => ['title.vector' => ['vector' => $vector, 'k' => 5]]],
],
'minimum_should_match' => 1,
],
];
}
/**
* 构建 fuzzy(多字段模糊)查询
*
* 按查询语种路由字段,避免巴利文被中文分词器切碎、中文被巴利文 analyzer 误伤:
* - 含汉字 → 只搜 bold_* + *.text.zh + summary.text
* - 巴利文(拉丁字母)→ 走三层排序(见 buildPaliFuzzyQuery)+ bold_* 提权
*
* @param string $query 搜索关键词
* @param array|null $weights 巴利文各词的罕见度权重,null 表示都按 1
* @return array OpenSearch DSL query 片段
*/
protected function buildFuzzyQuery(string $query, ?array $weights = null): array
{
if ($this->isCjkQuery($query)) {
return [
'multi_match' => [
'query' => $query,
'fields' => [
'bold_single^50',
'bold_multi^10',
'title.text.zh^3',
'summary.text^2',
'content.text.zh^1',
],
'type' => 'best_fields',
],
];
}
return $this->buildPaliFuzzyQuery($query, 1.0, $weights);
}
/**
* 巴利文 fuzzy 的计分表(全部是固定分,逐项累加)
*
* 一个文档的得分 = 命中的每一项相加,于是「部分符合」自然逐级降分:
* - phrase 整句按原顺序、拼写完全相同(≥2 个词才有)
* - bigram 相邻两词按原顺序、拼写相同,每对一次(≥3 个词才有,2 个词时与 phrase 重合)
* - exact 单个原始词拼写相同,每词一次
* - synonym 单个原始词经同义词可替代,每词一次(拼写相同的词也会命中这一项)
*
* 以 3 个词为例:整句原样 66 > 全部拼写相同但乱序 36 > 两词相邻拼写相同 + 一词同义 31
* > 两词拼写相同不相邻 24 > 三词全靠同义词 6 > 一词同义 2。
*
* @var array{phrase: int, bigram: int, exact: int, synonym: int}
*/
private const PALI_FUZZY_SCORES = [
'phrase' => 20,
'bigram' => 5,
'exact' => 10,
'synonym' => 2,
];
/**
* definition 的计分表(黑体字部分,每个原始词各算一次)
*
* 主分不分词形:同义词词典的一行就是一个词的全部变格变位,所以「任一词形」走同义词,
* 与 v2 search-pali-wbw 把全部词形一视同仁的口径一致;原拼写只给少量加分,排在同档前面。
*
* - single 该词(任一词形)是单独黑体字
* - single_exact 再加:拼写也相同
* - multi 该词(任一词形)出现在多词黑体短语里
* - multi_exact 再加:拼写也相同
* - multi_phrase 整句按原顺序、拼写相同出现在多词黑体短语里(≥2 个词才有)
*
* 每个词:单独黑体 ≥ 100 > 多词黑体 ≤ 35,于是命中的单独黑体字越多排越前。
* 正文得分按 DEFINITION_TEXT_SCALE 缩小后叠加,只用来给黑体同分的文档排先后。
*
* @var array{single: int, single_exact: int, multi: int, multi_exact: int, multi_phrase: int}
*/
private const DEFINITION_SCORES = [
'single' => 100,
'single_exact' => 10,
'multi' => 30,
'multi_exact' => 5,
'multi_phrase' => 30,
];
/**
* definition 模式里正文得分的缩放系数,保证正文再怎么命中也压不过一个黑体字
*/
private const DEFINITION_TEXT_SCALE = 0.1;
/**
* 同分时的 BM25 微调权重:BM25 一般 < 30,乘上后 < 0.3,小于计分表的最小档差
*/
private const TIE_BREAKER_BOOST = 0.01;
/**
* 罕见度权重的下限:再常见的词也保留这么多分量,不至于完全不计
*/
private const RARITY_WEIGHT_FLOOR = 0.2;
/**
* 罕见度权重的指数:IDF 取了对数,887 段与 211 段的词只差 0.82 : 1,拉不开;
* 归一后再取幂放大差距
*/
private const RARITY_EXPONENT = 4;
/**
* 每个词的文档频率在 RedisCache 里的保存期(秒)。语料按月更新,一天足够新
*/
private const DOC_FREQUENCY_TTL = 86400;
/**
* 多词查询里每个词的罕见度权重(类 BM25 的 IDF,按最罕见的词归一到 1)
*
* 按词计的固定分让每个词分量相同,常见词(如 bojjhaṅga,到处都是黑体)会把罕见的
* 目标词(dhammavicayasambojjhaṅga)挤出前排。乘上这个权重后,越罕见的词分越高。
* 文档频率按 title/content 的 .pali(带同义词,即全部词形)统计,一次 _msearch 取齐,
* 逐词缓存。取不到时返回全 1(等于不加权),只记日志不报错。
*
* @param list $words 参与计分的原始词(已去掉停用词)
* @return array 原始词 => 权重,范围 [RARITY_WEIGHT_FLOOR, 1]
*/
protected function wordRarityWeights(array $words): array
{
$uniqueWords = array_values(array_unique($words));
$flat = array_fill_keys($uniqueWords, 1.0);
if (count($uniqueWords) < 2) {
return $flat;
}
try {
$index = config('mint.opensearch.index');
$total = RedisCache::remember("/opensearch/{$index}/doc-count", self::DOC_FREQUENCY_TTL, fn () => $this->count());
$frequencies = $this->documentFrequencies($index, $uniqueWords);
} catch (Exception $e) {
Log::warning('opensearch word rarity weights failed', ['words' => $uniqueWords, 'error' => $e->getMessage()]);
return $flat;
}
$idf = [];
foreach ($uniqueWords as $word) {
$df = $frequencies[$word] ?? 0;
$idf[$word] = log(1 + ($total - $df + 0.5) / ($df + 0.5));
}
$maxIdf = max($idf);
if ($maxIdf <= 0) {
return $flat;
}
return array_map(
fn (float $value): float => max(self::RARITY_WEIGHT_FLOOR, ($value / $maxIdf) ** self::RARITY_EXPONENT),
$idf,
);
}
/**
* 每个词出现在多少个文档里(含同义词展开的全部词形),逐词缓存,未缓存的一次 _msearch 取齐
*
* @param string $index 索引名
* @param list $words 原始词
* @return array 原始词 => 文档数
*/
protected function documentFrequencies(string $index, array $words): array
{
$frequencies = [];
$missing = [];
foreach ($words as $word) {
$cached = RedisCache::get($this->documentFrequencyCacheKey($index, $word));
if ($cached === null) {
$missing[] = $word;
} else {
$frequencies[$word] = (int) $cached;
}
}
if ($missing === []) {
return $frequencies;
}
$body = [];
foreach ($missing as $word) {
$body[] = ['index' => $index];
$body[] = [
'size' => 0,
'track_total_hits' => true,
'query' => ['multi_match' => ['query' => $word, 'fields' => ['title.text.pali', 'content.text.pali']]],
];
}
$responses = $this->client->msearch(['body' => $body])['responses'] ?? [];
foreach ($missing as $i => $word) {
$df = (int) ($responses[$i]['hits']['total']['value'] ?? 0);
$frequencies[$word] = $df;
RedisCache::put($this->documentFrequencyCacheKey($index, $word), $df, self::DOC_FREQUENCY_TTL);
}
return $frequencies;
}
/**
* @param string $index 索引名
* @param string $word 原始词
* @return string 文档频率的缓存 key
*/
protected function documentFrequencyCacheKey(string $index, string $word): string
{
return "/opensearch/{$index}/word-df/".md5(mb_strtolower($word));
}
/**
* 巴利文停用词(小写;比较前去掉标点)
*
* 取自 pg 全文检索的 pali.stop(api-v8/public/app/fts/pali.stop),另加 vā。
* 停用词只参与词序打分(整句、相邻两词),不单独计分、不单独召回,见 buildPaliFuzzyQuery。
*
* @var list
*/
private const PALI_STOPWORDS = ['na', 'no', 'ca', 'va', 'vā', 'hi', 'pana', 'iva', 'eva'];
/**
* 判断原始词是否巴利文停用词
*
* @param string $word 原始词(可带标点,如「ca,」「‘‘na」)
* @return bool 是停用词返回 true
*/
protected function isPaliStopword(string $word): bool
{
$normalized = mb_strtolower(preg_replace('/[^\p{L}\p{M}]/u', '', $word) ?? '');
return in_array($normalized, self::PALI_STOPWORDS, true);
}
/**
* 取出用于单独计分的词:去掉停用词;全是停用词时原样返回,避免查不到东西
*
* @param list $words 原始词
* @return list
*/
protected function scoringWords(array $words): array
{
$contentWords = array_values(array_filter(
$words,
fn (string $word): bool => ! $this->isPaliStopword($word),
));
return $contentWords === [] ? $words : $contentWords;
}
/**
* 把查询切成原始词
*
* 按空白切分,并丢掉不含字母的片段(如「–」「.」):它们经分析器后没有 token,
* 留着会让按词计分的子句永远不命中。
*
* @param string $query 搜索关键词
* @return list
*/
protected function splitQueryWords(string $query): array
{
$words = preg_split('/\s+/u', trim($query)) ?: [];
return array_values(array_filter(
$words,
fn (string $word): bool => preg_match('/\p{L}/u', $word) === 1,
));
}
/**
* 包一层 constant_score:命中即得固定分,不受 BM25 字段长度归一化影响
*
* @param array $filter 过滤子句
* @param float $boost 命中时的固定分
* @return array{constant_score: array{filter: array, boost: float}}
*/
protected function constantScore(array $filter, float $boost): array
{
return ['constant_score' => ['filter' => $filter, 'boost' => $boost]];
}
/**
* 构建巴利文查询:词序符合、拼写相同得高分,部分符合逐级降分
*
* 命中按「原始词」(空格切分)计,不按同义词展开后的 token 计——每个原始词一个
* constant_score 子句,命中几个同义词都只算一次。计分表见 PALI_FUZZY_SCORES。
* 拼写相同打在 .exact 子字段(markdown_clean,无同义词),同义词打在 .pali。
* 最后加一条极小权重的 BM25 子句,只用来给同分文档排先后。
*
* 停用词(PALI_STOPWORDS)只出现在整句、相邻两词的子句里,保住词序判断;
* 不生成单词子句,也不进 BM25 子句,于是只命中「ca」的文档不会被召回。
* 两个停用词组成的相邻词对同理跳过。查询全是停用词时按普通词处理。
*
* 单词子句(拼写、同义词)乘上该词的罕见度权重(见 wordRarityWeights),
* 常见词分量小、罕见词分量大;词序子句(整句、相邻两词)不乘。
*
* @param string $query 搜索关键词
* @param float $scale 整体缩放(definition 模式用它把正文得分压到黑体字之下)
* @param array|null $weights 原始词 => 罕见度权重,null 表示都按 1
* @return array OpenSearch DSL query 片段
*/
protected function buildPaliFuzzyQuery(string $query, float $scale = 1.0, ?array $weights = null): array
{
$words = $this->splitQueryWords($query);
$wordCount = count($words);
if ($wordCount === 0) {
return ['match_all' => new \stdClass];
}
$scoringWords = $this->scoringWords($words);
$allStopwords = array_all($words, fn (string $word): bool => $this->isPaliStopword($word));
$exactFields = ['title.text.pali.exact', 'content.text.pali.exact'];
$fuzzyFields = ['title.text.pali', 'content.text.pali'];
$scores = array_map(fn (int $score): float => $score * $scale, self::PALI_FUZZY_SCORES);
$should = [];
// 词序:整句原样
if ($wordCount >= 2) {
$should[] = $this->constantScore([
'multi_match' => [
'query' => implode(' ', $words),
'type' => 'phrase',
'fields' => $exactFields,
],
], $scores['phrase']);
}
// 词序:相邻两词原样(部分顺序符合)
if ($wordCount >= 3) {
for ($i = 0; $i < $wordCount - 1; $i++) {
$bothStopwords = $this->isPaliStopword($words[$i]) && $this->isPaliStopword($words[$i + 1]);
if ($bothStopwords && ! $allStopwords) {
continue;
}
$should[] = $this->constantScore([
'multi_match' => [
'query' => $words[$i].' '.$words[$i + 1],
'type' => 'phrase',
'fields' => $exactFields,
],
], $scores['bigram']);
}
}
// 拼写:每个词各算一次(停用词不算),按罕见度加权
foreach ($scoringWords as $word) {
$weight = $weights[$word] ?? 1.0;
$should[] = $this->constantScore([
'multi_match' => [
'query' => $word,
'fields' => $exactFields,
'operator' => 'and',
],
], $scores['exact'] * $weight);
$should[] = $this->constantScore([
'multi_match' => [
'query' => $word,
'fields' => $fuzzyFields,
],
], $scores['synonym'] * $weight);
}
// 同分文档按 BM25 排先后
$should[] = [
'match' => [
'content.text.pali' => [
'query' => implode(' ', $scoringWords),
'boost' => self::TIE_BREAKER_BOOST * $scale,
],
],
];
return [
'bool' => [
'should' => $should,
],
];
}
/**
* 判断查询是否含中文(CJK 统一表意文字)
*
* @param string $query 搜索关键词
* @return bool 含汉字返回 true,否则按巴利文(拉丁字母)处理
*/
protected function isCjkQuery(string $query): bool
{
return (bool) preg_match('/[\x{4E00}-\x{9FFF}\x{3400}-\x{4DBF}\x{F900}-\x{FAFF}]/u', $query);
}
/**
* 构建 definition(定义查询)
*
* 义注里的黑体字通常就是被解释的词,所以黑体字命中决定排序,正文只做次级排序:
* - 每个原始词单独算:命中的单独黑体字(bold_single)越多分越高
* - 单独黑体字 > 多词黑体短语(bold_multi)里的词
* - 词形不分高下(走同义词,即该词的全部变格变位),原拼写只在同档内加少量分:
* bold_* 建索引用 standard,查询时显式指定 analyzer=standard 即是「拼写相同」,
* 不指定则走带同义词的 pali_query_analyzer
* - 多词黑体也按任一词形算:bold_* 取自 wbw_templates.real(已拆连音),
* 正文里粘了连音(如 vitakkanti)匹配不到的词,往往靠它召回
* - 用 match_phrase 而不是 match:中文在 standard 下按字切分,phrase 要求字连续,
* 避免一个字重合就命中
* - 每个词的黑体分乘上它的罕见度权重(见 wordRarityWeights):常见词(bojjhaṅga)
* 不再与罕见的目标词(dhammavicayasambojjhaṅga)同分,挤占前排
* 计分表见 DEFINITION_SCORES;正文部分复用 fuzzy 的计分,乘 DEFINITION_TEXT_SCALE。
*
* @param string $query 搜索关键词
* @param array|null $weights 原始词 => 罕见度权重,null 表示都按 1
* @return array OpenSearch DSL query 片段
*/
protected function buildDefinitionQuery(string $query, ?array $weights = null): array
{
$words = $this->splitQueryWords($query);
if ($words === []) {
return ['match_all' => new \stdClass];
}
$should = [];
// 停用词不单独算黑体分(整句的 multi_phrase 仍带上它们)
foreach ($this->scoringWords($words) as $word) {
$weight = $weights[$word] ?? 1.0;
foreach (['single' => 'bold_single', 'multi' => 'bold_multi'] as $tier => $field) {
$should[] = $this->constantScore(
['match_phrase' => [$field => $word]],
self::DEFINITION_SCORES[$tier] * $weight,
);
$should[] = $this->constantScore(
['match_phrase' => [$field => ['query' => $word, 'analyzer' => 'standard']]],
self::DEFINITION_SCORES[$tier.'_exact'] * $weight,
);
}
}
if (count($words) >= 2) {
$should[] = $this->constantScore(
['match_phrase' => ['bold_multi' => ['query' => implode(' ', $words), 'analyzer' => 'standard']]],
self::DEFINITION_SCORES['multi_phrase'],
);
}
// 正文:黑体同分时的次级排序,也保留没有黑体的文档(偈颂等)的召回
if ($this->isCjkQuery($query)) {
$should[] = [
'multi_match' => [
'query' => $query,
'fields' => ['title.text.zh^3', 'content.text.zh'],
'boost' => self::DEFINITION_TEXT_SCALE,
],
];
} else {
$should[] = $this->buildPaliFuzzyQuery($query, self::DEFINITION_TEXT_SCALE, $weights);
}
return [
'bool' => [
'should' => $should,
],
];
}
/**
* 构建 hybrid(模糊 + 语义混合)查询
*
* 使用 bool should 将 fuzzy(constant_score 包裹)与三路 knn 向量查询合并,
* 权重比例由 $weights['hybrid']['fuzzy_ratio'] 和 'semantic_ratio' 控制。
* title.vector 的语义权重略高(×1.2),以提升标题匹配的排名。
*
* @param string $query 搜索关键词
* @return array OpenSearch DSL query 片段
*
* @throws Exception embedding 调用失败时抛出
*/
protected function buildHybridQuery(string $query): array
{
$fuzzyFields = [];
foreach ($this->weights['hybrid'] as $field => $weight) {
if (in_array($field, ['fuzzy_ratio', 'semantic_ratio'])) {
continue;
}
$fuzzyFields[] = $field.'^'.$weight;
}
$fuzzyPart = [
'multi_match' => [
'query' => $query,
'fields' => $fuzzyFields,
'type' => 'best_fields',
],
];
$vector = $this->embedText($query);
$fuzzyRatio = $this->weights['hybrid']['fuzzy_ratio'];
$semanticRatio = $this->weights['hybrid']['semantic_ratio'];
return [
'bool' => [
'should' => [
[
'constant_score' => [
'filter' => $fuzzyPart,
'boost' => $fuzzyRatio,
],
],
[
'knn' => [
'content.vector' => [
'vector' => $vector,
'k' => 20,
'boost' => $semanticRatio * 1.0,
],
],
],
[
'knn' => [
'summary.vector' => [
'vector' => $vector,
'k' => 10,
'boost' => $semanticRatio * 0.8,
],
],
],
[
'knn' => [
'title.vector' => [
'vector' => $vector,
'k' => 5,
'boost' => $semanticRatio * 1.2, // title 权重略高
],
],
],
],
],
];
}
/**
* 调用 OpenAI Embedding API 将文本转为向量
*
* 使用 Redis 缓存(TTL 7 天),相同文本不会重复请求 API,
* 缓存 key 格式见 embeddingCacheKey()。
*
* @param string $text 输入文本
* @return array 1536 维 float 向量
*
* @throws Exception 未设置 OPENAI_API_KEY 或 API 返回异常时抛出
*/
protected function embedText(string $text): array
{
if (! $this->openaiApiKey) {
throw new Exception('请在 .env 设置 OPENAI_API_KEY');
}
return RedisCache::remember($this->embeddingCacheKey($text), now()->addDays(7), function () use ($text) {
$response = $this->http->post('embeddings', [
'headers' => [
'Authorization' => 'Bearer '.$this->openaiApiKey,
'Content-Type' => 'application/json',
],
'json' => [
'model' => 'text-embedding-3-small',
'input' => $text,
],
]);
$json = json_decode((string) $response->getBody(), true);
if (empty($json['data'][0]['embedding'])) {
throw new Exception('OpenAI embedding 返回异常: '.json_encode($json));
}
return $json['data'][0]['embedding'];
});
}
/**
* embedding 缓存键:"embedding/v{版本}/{md5(text)}"。
*
* 版本号存在 EMBEDDING_CACHE_VERSION_KEY,clearAllEmbeddingCache() 把它加一,
* 旧版本的键就再也不会被读到,7 天后自然过期。原来用 Cache::tags(),
* tag 索引是另外的 key,在 Redis Cluster 上跨 slot。
*/
private function embeddingCacheKey(string $text): string
{
$version = RedisCache::get(self::EMBEDDING_CACHE_VERSION_KEY, 0);
return "embedding/v{$version}/".md5($text);
}
/**
* 清除指定文本的 embedding 缓存
*
* @param string $text 原始文本(与调用 embedText 时一致)
* @return bool 缓存是否成功删除
*
* @example
* $service->clearEmbeddingCache('sabbe dhammā anattā');
*/
public function clearEmbeddingCache(string $text): bool
{
return RedisCache::forget($this->embeddingCacheKey($text));
}
/**
* 让所有 embedding 缓存失效
*
* 不逐条删除,而是把版本号加一;旧键 7 天内自然过期。
*
* @example
* $service->clearAllEmbeddingCache();
*/
public function clearAllEmbeddingCache(): void
{
RedisCache::increment(self::EMBEDDING_CACHE_VERSION_KEY);
}
/**
* 自动建议(Completion Suggest)
*
* 基于 completion 字段实现前缀补全,支持同时查询多个语言字段。
* 结果按 _score 降序排序,跨字段去重。
*
* 可用字段标识符($fields 参数):
* - 'title_pali' → title.suggest.pali
* - 'title_zh' → title.suggest.zh
* - 'content_pali' → content.suggest.pali
* - 'content_zh' → content.suggest.zh
*
* @param string $query 查询前缀文本
* @param array|string|null $fields 要查询的字段标识符,null 表示全部字段
* @param string|null $language 可选的语言过滤(term query)
* @param int $limit 每个字段返回的建议数量,默认 10
* @return array 建议结果列表,每项包含:
* text, source(字段标识符), score, doc_id, doc_source
*
* @throws \InvalidArgumentException $fields 中含无效字段标识符时抛出
*
* @example
* // 查询所有字段
* $service->suggest('nibb');
*
* // 只查询巴利文标题建议
* $service->suggest('nibb', 'title_pali');
*
* // 查询多个字段,限制语言
* $service->suggest('涅', ['title_zh', 'content_zh'], 'zh', 5);
*/
public function suggest(
string $query,
$fields = null,
?string $language = null,
int $limit = 10
): array {
// 字段标识符 → OpenSearch completion 字段路径
$fieldMap = [
'title_pali' => 'title.suggest.pali',
'title_zh' => 'title.suggest.zh',
'content_pali' => 'content.suggest.pali',
'content_zh' => 'content.suggest.zh',
];
// 处理字段参数
if ($fields === null) {
$searchFields = array_keys($fieldMap);
} elseif (is_string($fields)) {
$searchFields = [$fields];
} else {
$searchFields = $fields;
}
// 过滤无效字段
$searchFields = array_values(array_filter(
$searchFields,
fn ($field) => isset($fieldMap[$field])
));
if (empty($searchFields)) {
throw new \InvalidArgumentException('Invalid fields specified for suggestion');
}
// 构建 suggest DSL
$suggests = [];
foreach ($searchFields as $field) {
$suggests[$field.'_suggest'] = [
'prefix' => $query,
'completion' => [
'field' => $fieldMap[$field],
'size' => $limit,
'skip_duplicates' => true,
],
];
}
$dsl = ['suggest' => $suggests];
if ($language) {
$dsl['query'] = ['term' => ['language' => $language]];
}
$response = $this->client->search([
'index' => config('mint.opensearch.index'),
'body' => $dsl,
]);
// 整理结果,附加来源字段
$results = [];
foreach ($searchFields as $field) {
$options = $response['suggest'][$field.'_suggest'][0]['options'] ?? [];
foreach ($options as $opt) {
$results[] = [
'text' => $opt['text'] ?? '',
'source' => $field,
'score' => $opt['_score'] ?? 0,
'doc_id' => $opt['_id'] ?? null,
'doc_source' => $opt['_source'] ?? null,
];
}
}
// 按分数降序排序
usort($results, fn ($a, $b) => $b['score'] <=> $a['score']);
return $results;
}
/**
* 按文档 ID 获取单条完整文档(包含 content.display)
*
* @param string $id 文档 ID,例如 "term_{guid}"
* @return array OpenSearch 原始响应
*/
public function get(string $id): array
{
return $this->client->get([
'index' => config('mint.opensearch.index'),
'id' => $id,
]);
}
/**
* 校验 pali_synonyms 同义词词典是否对指定词生效
*
* 通过 OpenSearch _analyze API,用当前索引的 pali_query_analyzer
* 对输入文本做实时分析,返回展开后的全部 token(含原词与同义词)。
* 可用于快速确认 analysis/pali_synonyms.txt 中的某一行是否已生效
* (例如 dhamma,dharma,法 => dhamma 是否真的展开出 dharma、法)。
*
* @param string $text 待检测的巴利词,例如 "dhamma"
* @return array 分析器输出的 token 文本数组(已去重,保留原始顺序)
*
* @throws Exception 索引不存在或 OpenSearch 调用失败时抛出
*
* @example
* $service->pali_query_health_check('dhamma');
* // => ['dhamma', 'dharma', '法']
*/
public function pali_query_health_check(string $text): array
{
$index = config('mint.opensearch.index');
if (! $this->client->indices()->exists(['index' => $index])) {
throw new Exception("Index [$index] does not exist.");
}
$response = $this->client->indices()->analyze([
'index' => $index,
'body' => [
'analyzer' => 'pali_query_analyzer',
'text' => $text,
],
]);
$tokens = array_map(
fn ($token) => $token['token'] ?? '',
$response['tokens'] ?? []
);
return array_values(array_unique(array_filter($tokens, fn ($t) => $t !== '')));
}
}