参考文献校对，本地大模型提示词记录
This commit is contained in:
609
application/common/service/ReferenceRelevanceLlmService.php
Normal file
609
application/common/service/ReferenceRelevanceLlmService.php
Normal file
@@ -0,0 +1,609 @@
|
||||
<?php
|
||||
|
||||
namespace app\common\service;
|
||||
|
||||
use think\Env;
|
||||
|
||||
/**
|
||||
* 参考文献「主题相关性」LLM 校对（独立于支撑力度校对 LLMService）
|
||||
*/
|
||||
class ReferenceRelevanceLlmService
|
||||
{
|
||||
private $url;
|
||||
private $model;
|
||||
private $apiKey;
|
||||
private $timeout;
|
||||
private $lastPostError = '';
|
||||
private $maxSectionChars;
|
||||
private $maxLocalContextChars;
|
||||
private $maxReferChars;
|
||||
private $maxAbstractChars;
|
||||
|
||||
public function __construct()
|
||||
{
|
||||
$this->url = trim((string)Env::get('promotion.promotion_llm_url', ''));
|
||||
$this->model = trim((string)Env::get('promotion.promotion_llm_model', ''));
|
||||
$this->apiKey = trim((string)Env::get('promotion.promotion_llm_api_key', ''));
|
||||
$this->timeout = max(180, intval(Env::get('promotion.promotion_llm_timeout', 180)));
|
||||
// 控制发送给 LLM 的上下文长度，降低单次推理耗时（可通过 env 覆盖）
|
||||
$this->maxSectionChars = max(1500, intval(Env::get('promotion.relevance_llm_max_section_chars', 4500)));
|
||||
$this->maxLocalContextChars = max(600, intval(Env::get('promotion.relevance_llm_max_local_context_chars', 1800)));
|
||||
$this->maxReferChars = max(1500, intval(Env::get('promotion.relevance_llm_max_refer_chars', 3500)));
|
||||
$this->maxAbstractChars = max(1500, intval(Env::get('promotion.relevance_llm_max_abstract_chars', 3500)));
|
||||
}
|
||||
|
||||
/**
|
||||
* @return array{results:array,request_failed?:bool,reason?:string}
|
||||
*/
|
||||
public function checkRelevance($sectionText, $localContext, $referText, $abstractText = '', $citeGroupRefs = '')
|
||||
{
|
||||
$fallback = [
|
||||
'results' => [],
|
||||
'request_failed' => true,
|
||||
'reason' => 'LLM not configured or request failed',
|
||||
];
|
||||
if ($this->url === '' || $this->model === '') {
|
||||
return $fallback;
|
||||
}
|
||||
|
||||
$sectionText = trim((string)$sectionText);
|
||||
$localContext = trim((string)$localContext);
|
||||
$referText = trim((string)$referText);
|
||||
$abstractText = trim((string)$abstractText);
|
||||
if ($sectionText === '' || $referText === '') {
|
||||
return ['results' => [], 'reason' => 'Empty section or reference text'];
|
||||
}
|
||||
|
||||
if (mb_strlen($sectionText) > $this->maxSectionChars) {
|
||||
$sectionText = mb_substr($sectionText, 0, $this->maxSectionChars);
|
||||
}
|
||||
if (mb_strlen($localContext) > $this->maxLocalContextChars) {
|
||||
$localContext = mb_substr($localContext, 0, $this->maxLocalContextChars);
|
||||
}
|
||||
if (mb_strlen($referText) > $this->maxReferChars) {
|
||||
$referText = mb_substr($referText, 0, $this->maxReferChars);
|
||||
}
|
||||
if (mb_strlen($abstractText) > $this->maxAbstractChars) {
|
||||
$abstractText = mb_substr($abstractText, 0, $this->maxAbstractChars);
|
||||
}
|
||||
|
||||
$payload = [
|
||||
'model' => $this->model,
|
||||
'temperature' => 0,
|
||||
'messages' => [
|
||||
['role' => 'system', 'content' => $this->buildSystemPrompt()],
|
||||
['role' => 'user', 'content' => $this->buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs)],
|
||||
],
|
||||
];
|
||||
|
||||
$content = $this->postChat($payload);
|
||||
if ($content === null) {
|
||||
$reason = $this->lastPostError !== '' ? $this->lastPostError : 'LLM request failed';
|
||||
return array_merge($fallback, ['reason' => $reason]);
|
||||
}
|
||||
|
||||
$parsed = $this->parseJson($content);
|
||||
if ($parsed === null) {
|
||||
return array_merge($fallback, ['reason' => 'LLM response JSON parse failed']);
|
||||
}
|
||||
|
||||
$results = $this->normalizeResults($parsed, $citeGroupRefs, $localContext, $referText, $abstractText);
|
||||
if (empty($results)) {
|
||||
return array_merge($fallback, ['reason' => 'LLM returned empty or invalid results']);
|
||||
}
|
||||
|
||||
return ['results' => $results];
|
||||
}
|
||||
|
||||
private function buildSystemPrompt()
|
||||
{
|
||||
return <<<'PROMPT'
|
||||
你是一名护理、医学、生物医学与科研期刊的资深学术编辑，正在执行「参考文献主题相关性校对」。
|
||||
|
||||
你的任务：判断【引用位置正文表述】与【对应编号参考文献】在主题、研究对象、疾病/场景/结局方向上是否相关，能否作为该处引用的合理来源。
|
||||
|
||||
注意：这是「相关性」校对，侧重引用处具体 claim 与文献内容是否匹配；**不是**判断「是否同一疾病/同一领域」。
|
||||
|
||||
==================================================
|
||||
【零、最硬规则（违反则输出无效）】
|
||||
1. **单条 relevance_score 只评价该编号文献单独**与引用处的关系；不得因联合组整体合理而抬高弱相关文献的单条分。
|
||||
2. **禁止「同病高分」**：正文与文献都涉及 CRC，不等于单条可给 0.85~0.92。
|
||||
**但若引用处 claim 本身就是机制/通路/异质性/耐药/治疗挑战**，且**研究主语一致**（同一疾病/同一化合物/同一干预对象），文献（含摘要/清洗内容）讨论同病多通路、遗传改变、耐药等，应给 **0.65~0.78**，不得误降到 0.45。
|
||||
**主语不一致时仍适用本条禁止高分**：引用处主语为化合物 X，文献却是其他植物/提取物/计算预测，即使提到 X 或相同通路名，也不得因此给 0.78+。
|
||||
3. 引用处若为**流行病学/负担类 claim**（most common、incidence、mortality、burden、全球高发等）：
|
||||
- 机制研究、分子通路、细胞增殖/迁移、血管生成等**原始研究** → 单条通常 **0.45 或更低**，`is_relevant=0`，`minimal_relevance`
|
||||
- 不得因摘要提到 colorectal cancer 就给 0.92
|
||||
- 仅当文献为流行病学综述/公共卫生研究，或明确讨论发病率、死亡率、疾病负担时，单条才可 **0.85~0.92**
|
||||
4. **联合分写在 combined_relevance_score**，与单条分必须可分离（例如 [1,2] 时文献1=0.45、文献2=0.92、联合=0.92）。
|
||||
5. **「来源/化学分类」型句子**（naturally occurring、pentacyclic triterpenoid、found in fruits/vegetables/medicinal plants、并列举具体植物学名）：
|
||||
- 先判文献类型：来源综述 / 生物活性综述 最适合；**抗癌治疗综述**对「来源分布」claim 通常仅 **0.65**
|
||||
- 单篇可差异化打分（如 0.92 / 0.92 / 0.65），**不得**因联合而三篇都给高分
|
||||
- 若原句含**具体列举项**（如多个植物学名），而材料未逐一核实全部学名，联合分通常 **≤0.85**（不得给 0.98）
|
||||
6. **多要素综括句**（一句同时塞入：药学/研究兴趣 + 大量前临床研究 + 多种活性[抗炎/抗氧化/抗癌等] + 多个癌种/对象列举）：
|
||||
- 单篇即使是综述，通常仅 partially_related ~ near-direct（**0.78~0.86**），**不轻易给 0.92**（单篇难逐项覆盖全部要素）
|
||||
- **联合分是整句覆盖度评估，可低于最高单条分**：若整句要素需多篇拼合、且含作者整合概括，联合通常 **0.72~0.78（partially_related）**，不给 0.85+
|
||||
7. **联合分不是「取最高单条分」**：当各单篇都只覆盖整句一部分、需互补拼合时，联合分应反映「整句作为一个整体被支撑的完整度」，**允许低于任何一篇单条分**。
|
||||
8. **主语/研究对象层级必须对齐**：引用处主语为某化合物/分子（如「X has been demonstrated…」）时，文献核心对象须为 **X 本身**或以 X 为核心的实验/综述。
|
||||
- **植物提取物/混合物**研究、**其他物种/其他植物**的计算预测、成分表中顺带出现 X → 通常 **0.45 或更低**
|
||||
- **关键：提取物即使 X 含量很高（如 50%+）且显示了抗癌/凋亡活性，活性归因于提取物整体而非 X 单体单独验证 → 仍属 weakly_related（≤0.45）、`minimal_relevance`**，不得评为 supplementary_relevance/0.78
|
||||
- 只有当文献**针对 X 单体单独做了验证**（X monomer 处理、X 单独剂量效应等）时，主语才算对齐，方可进入 0.65+
|
||||
- **不得**因摘要/讨论出现与引用句相同的通路名、凋亡、抗癌等词就给 0.78+
|
||||
9. **证据层级与 demonstrated / mechanistically**：
|
||||
- 本文实验结果或针对 X 的系统综述 > 计算预测/混合成分推测
|
||||
- **讨论（Discussion）转引他人关于 X 的机制总结 ≠ 该文自身证据**；据此最多 **0.45~0.65**，不得评为 highly_related
|
||||
- in silico / computational prediction 不足以支撑「has been demonstrated to mechanistically…」式强语气 claim 的高分
|
||||
10. **点名通路/功能结局须逐项核对**：原句逐条列举通路（如 PI3K/AKT、MAPK、NF-κB）或结局（增殖、凋亡、血管生成、炎症信号等）时，**每一项单独核对是否在本文证据中成立**（非仅背景提及）。
|
||||
- 讨论转述既往文献 ≠ 本文证明该项
|
||||
- 缺原句任一点名项（如 angiogenesis）→ 单条通常 **不得 0.78+**
|
||||
- **「覆盖部分结局」不足以进入 0.78**：原句点名了多条通路 + 多个结局，文献仅命中其中 1~2 个结局（如仅凋亡/增殖），且**点名通路在本文结果中全部缺失（仅讨论转引）**或主语层级不对 → 单条 **限 0.45（weakly_related / minimal_relevance）**，不得给 0.65~0.78
|
||||
- 仅同领域沾边 1–2 项、主语或机制层级不对 → **0.45**
|
||||
- **进入 0.65~0.78 的前提**：主语对齐（X 单体）+ 本文自身结果命中原句点名通路/结局的多数项；几乎全部明确对应 → **0.85+**
|
||||
|
||||
==================================================
|
||||
【一、必须先拆解 claim】
|
||||
从【本引用位置附近上下文】中提炼最小主张单元（Claim A, Claim B…），**不要**把整句笼统归为「大概讲抗癌」。例如：
|
||||
- **主语/研究对象**（化合物单体 vs 植物提取物 vs 其他物种；是否「X has been demonstrated」）
|
||||
- **证据语气与层级**（demonstrated / mechanistically vs predict / suggest；本文结果 vs 讨论转引）
|
||||
- 疾病流行病学（高发、死亡率）
|
||||
- **点名通路/分子机制**（PI3K/AKT、MAPK、NF-κB 等，须逐项）
|
||||
- **点名功能结局**（抑制增殖、凋亡、血管生成、炎症信号等，须逐项）
|
||||
- 治疗/干预现状
|
||||
- **化合物化学类别**（如 pentacyclic triterpenoid）
|
||||
- **天然来源分布**（fruits / vegetables / medicinal plants）
|
||||
- **具体列举项**（植物学名、药名、基因名等，须逐项核对）
|
||||
|
||||
==================================================
|
||||
【二、逐篇文献单独判断（每条 result 对应一个 reference_no）】
|
||||
对 cite_group_refs 中的每一篇文献，单独输出：
|
||||
- 该文献与引用处哪些 claim 主题相关、哪些不相关（含具体列举项是否覆盖）
|
||||
- 文献类型是否匹配引用用途（来源综述 / 生物活性综述 / 机制研究 / 流行病学综述 / 抗癌治疗综述等）
|
||||
- relevance_score：只能使用 0.98 / 0.92 / 0.85 / 0.78 / 0.65 / 0.45 / 0.25 / 0.15
|
||||
- relevance_level：highly_related | partially_related | weakly_related | unrelated
|
||||
- is_relevant：score>=0.65 为 1，否则 0
|
||||
- relevance_role：
|
||||
- primary_relevance：该文献是引用处主题的主要相关来源
|
||||
- supplementary_relevance：部分相关、补充性
|
||||
- minimal_relevance：仅边缘/背景沾边
|
||||
- no_meaningful_relevance：与引用处核心表述基本无关
|
||||
- reason：中文，须写明：①文献类型与**核心研究对象** ②**本文自身证据**覆盖了哪些 claim / 哪些未覆盖（含点名通路、功能结局、具体列举项；讨论转引须标明） ③**若仅为同主题不同 claim 或主语层级不对须明确写出** ④为何此分值而非更高或更低
|
||||
|
||||
主语/层级不对 → 单条 **0.45**，不得因讨论提及相同通路给 0.78：
|
||||
引用处 claim 为「化合物 X 经 PI3K/AKT 等机制 demonstrated…」，文献为其他植物提取物或计算预测、仅在讨论转引他人 X 机制 → 0.45，weakly_related，is_relevant=0。
|
||||
|
||||
机制文引用流行病学句 → 单条 **0.45**，不得 0.92：
|
||||
文献为 CRC 机制研究，引用处 claim 为全球高发/死亡率，文献无流行病学数据 → 0.45，minimal_relevance，is_relevant=0。
|
||||
|
||||
==================================================
|
||||
【三、联合引用 combined_*（同一 cite_group_refs 内各行必须一致）】
|
||||
当 cite_group_refs 为 "1,2" 等多篇时，除逐篇判断外，必须给出引用组整体结论：
|
||||
- 这些文献合起来，是否足以支撑/匹配该引用位置的整体表述？
|
||||
- combined_relevance_score：八档固定分值之一，**不是单条平均分**
|
||||
- 若一篇已强相关、其余仅弱补充，联合分可接近主相关文献，但**不必等于最高单条分**
|
||||
- 若原句含具体列举项（学名等）且材料未逐一核实，联合分通常 **0.85**，不给 0.98
|
||||
- 若核心 claim 无任何文献明确覆盖，联合分不能虚高
|
||||
- 多篇联合仍缺主语对齐、缺原句点名通路/结局、或主要靠讨论转引 → 联合分通常 **≤0.45~0.65**，不得因单篇讨论出现相同关键词给到 0.78+
|
||||
- combined_is_relevant：combined_relevance_score>=0.65 为 1
|
||||
- combined_relevance_level：与 combined 分数对应的等级
|
||||
- combined_reason：中文综合结论，说明各文献分工（如「文献2综述覆盖流行病学主 claim，文献1仅机制补充」）及最终分值理由
|
||||
|
||||
单条引用时：combined_* 与单条一致，combined_reason 可写「」。
|
||||
|
||||
==================================================
|
||||
【四、评分与等级对照】
|
||||
0.98 / 0.92 / 0.85 = highly_related
|
||||
文献直接支持整句主旨，大部分关键要素都在文中明确出现
|
||||
0.78 / 0.65 = partially_related
|
||||
文献只支撑其中一部分，或支撑方式偏间接
|
||||
0.45 = weakly_related
|
||||
只是同领域文献，但与句子事实对应很弱
|
||||
0.25 / 0.15 = unrelated
|
||||
基本不支撑该句
|
||||
≤0.15 = not_support
|
||||
不支撑
|
||||
|
||||
==================================================
|
||||
【五、输出 JSON（仅 JSON，无 markdown）】
|
||||
{
|
||||
"results": [
|
||||
{
|
||||
"reference_no": 1,
|
||||
"cite_group_refs": "1,2",
|
||||
"is_relevant": 0,
|
||||
"relevance_score": 0.45,
|
||||
"relevance_level": "weakly_related",
|
||||
"relevance_role": "minimal_relevance",
|
||||
"reason": "中文单条结论",
|
||||
"combined_is_relevant": 1,
|
||||
"combined_relevance_score": 0.92,
|
||||
"combined_relevance_level": "highly_related",
|
||||
"combined_reason": "中文联合结论"
|
||||
},
|
||||
{
|
||||
"reference_no": 2,
|
||||
"cite_group_refs": "1,2",
|
||||
...
|
||||
}
|
||||
]
|
||||
}
|
||||
PROMPT;
|
||||
}
|
||||
|
||||
private function buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs)
|
||||
{
|
||||
$parts = ["【正文节 t_article_main】\n" . $sectionText];
|
||||
if (trim((string)$citeGroupRefs) !== '') {
|
||||
$mode = strpos($citeGroupRefs, ',') !== false ? '联合引用' : '单独引用';
|
||||
$parts[] = "【引用文献组 cite_group_refs】{$citeGroupRefs}（{$mode}）";
|
||||
}
|
||||
if ($localContext !== '') {
|
||||
$parts[] = "【本引用位置附近上下文（优先据此拆解 claim）】\n" . $localContext;
|
||||
}
|
||||
$parts[] = "【参考文献书目（按编号）】\n" . $referText;
|
||||
if ($abstractText !== '') {
|
||||
$parts[] = "【文献摘要/清洗后内容（Europe PMC·PubMed·Crossref·PDF）】\n" . $abstractText;
|
||||
}
|
||||
$parts[] = '请先拆解最小主张单元（主语层级、证据来源、点名通路/结局逐项核对），判断每篇文献类型与**本文自身证据**，再**逐篇独立**给出单条 relevance_score（讨论转引、提取物/计算预测不得抬高；弱相关文献不得因联合而高分），最后给出 combined_*。仅输出 results 数组 JSON。';
|
||||
|
||||
return implode("\n\n", $parts);
|
||||
}
|
||||
|
||||
private function normalizeResults(array $parsed, $defaultCiteGroupRefs, $localContext = '', $referText = '', $abstractText = '')
|
||||
{
|
||||
$rows = [];
|
||||
if (isset($parsed['results']) && is_array($parsed['results'])) {
|
||||
$rows = $parsed['results'];
|
||||
} elseif (isset($parsed['reference_no']) || isset($parsed['relevance_score'])) {
|
||||
$rows = [$parsed];
|
||||
}
|
||||
|
||||
$bands = $this->getScoreBands();
|
||||
$localContext = trim((string)$localContext);
|
||||
$referText = trim((string)$referText);
|
||||
$abstractText = trim((string)$abstractText);
|
||||
|
||||
$out = [];
|
||||
foreach ($rows as $item) {
|
||||
if (!is_array($item)) {
|
||||
continue;
|
||||
}
|
||||
$refNo = intval(isset($item['reference_no']) ? $item['reference_no'] : 0);
|
||||
if ($refNo <= 0) {
|
||||
continue;
|
||||
}
|
||||
|
||||
$score = $this->snapScore(floatval(isset($item['relevance_score']) ? $item['relevance_score'] : 0), $bands);
|
||||
$isRelevant = $score >= 0.65 - 0.001;
|
||||
if (array_key_exists('is_relevant', $item)) {
|
||||
$isRelevant = $this->boolVal($item['is_relevant']);
|
||||
}
|
||||
|
||||
$level = $this->levelFromScore($score, isset($item['relevance_level']) ? $item['relevance_level'] : '');
|
||||
$role = $this->normalizeRelevanceRole(isset($item['relevance_role']) ? $item['relevance_role'] : '');
|
||||
$reason = $this->cleanReason(isset($item['reason']) ? $item['reason'] : '');
|
||||
|
||||
list($score, $level, $isRelevant, $role) = $this->enforceSingleReferenceConsistency(
|
||||
$score,
|
||||
$level,
|
||||
$isRelevant,
|
||||
$role,
|
||||
$bands
|
||||
);
|
||||
|
||||
$combinedScore = $this->snapScore(
|
||||
floatval(isset($item['combined_relevance_score']) ? $item['combined_relevance_score'] : $score),
|
||||
$bands
|
||||
);
|
||||
$combinedRelevant = $combinedScore >= 0.65 - 0.001;
|
||||
if (array_key_exists('combined_is_relevant', $item)) {
|
||||
$combinedRelevant = $this->boolVal($item['combined_is_relevant']);
|
||||
}
|
||||
|
||||
$combinedLevel = $this->levelFromScore(
|
||||
$combinedScore,
|
||||
isset($item['combined_relevance_level']) ? $item['combined_relevance_level'] : ''
|
||||
);
|
||||
list($combinedScore, $combinedLevel, $combinedRelevant) = $this->enforceCombinedConsistency(
|
||||
$combinedScore,
|
||||
$combinedLevel,
|
||||
$combinedRelevant,
|
||||
$bands
|
||||
);
|
||||
|
||||
$citeGroupRefs = trim((string)(isset($item['cite_group_refs']) ? $item['cite_group_refs'] : $defaultCiteGroupRefs));
|
||||
if ($citeGroupRefs === '' && $defaultCiteGroupRefs !== '') {
|
||||
$citeGroupRefs = trim((string)$defaultCiteGroupRefs);
|
||||
}
|
||||
|
||||
$out[] = [
|
||||
'reference_no' => $refNo,
|
||||
'cite_group_refs' => $citeGroupRefs,
|
||||
'is_relevant' => $isRelevant ? 1 : 0,
|
||||
'relevance_score' => $score,
|
||||
'relevance_level' => $level,
|
||||
'relevance_role' => $role,
|
||||
'reason' => $reason,
|
||||
'combined_is_relevant' => $combinedRelevant ? 1 : 0,
|
||||
'combined_relevance_score' => $combinedScore,
|
||||
'combined_relevance_level' => $combinedLevel,
|
||||
'combined_reason' => $this->cleanReason(isset($item['combined_reason']) ? $item['combined_reason'] : ''),
|
||||
];
|
||||
}
|
||||
|
||||
$out = $this->syncCombinedFieldsAcrossGroup($out);
|
||||
|
||||
return $out;
|
||||
}
|
||||
|
||||
private function enforceSingleReferenceConsistency($score, $level, $isRelevant, $role, array $bands)
|
||||
{
|
||||
$score = floatval($score);
|
||||
if ($role === 'no_meaningful_relevance') {
|
||||
if ($score > 0.25) {
|
||||
$score = 0.25;
|
||||
}
|
||||
$level = 'unrelated';
|
||||
$isRelevant = false;
|
||||
} elseif ($role === 'minimal_relevance') {
|
||||
if ($score > 0.45) {
|
||||
$score = 0.45;
|
||||
}
|
||||
$level = 'weakly_related';
|
||||
$isRelevant = false;
|
||||
} elseif ($role === 'supplementary_relevance') {
|
||||
if ($score > 0.78) {
|
||||
$score = 0.78;
|
||||
}
|
||||
$level = $this->levelFromScore($score, $level);
|
||||
} elseif ($role === 'primary_relevance') {
|
||||
if ($score < 0.85) {
|
||||
$score = 0.85;
|
||||
}
|
||||
$isRelevant = true;
|
||||
$level = $this->levelFromScore($score, $level);
|
||||
}
|
||||
|
||||
if ($level === 'weakly_related' && $score > 0.45) {
|
||||
$score = 0.45;
|
||||
$isRelevant = false;
|
||||
} elseif ($level === 'unrelated' && $score > 0.25) {
|
||||
$score = 0.25;
|
||||
$isRelevant = false;
|
||||
} elseif ($level === 'highly_related' && $score < 0.85) {
|
||||
$score = 0.85;
|
||||
$isRelevant = true;
|
||||
} elseif ($level === 'partially_related') {
|
||||
if ($score > 0.78) {
|
||||
$score = 0.78;
|
||||
}
|
||||
if ($score < 0.65) {
|
||||
$score = 0.65;
|
||||
}
|
||||
$isRelevant = true;
|
||||
}
|
||||
|
||||
if (!$isRelevant && $score >= 0.65) {
|
||||
$score = 0.45;
|
||||
$level = 'weakly_related';
|
||||
}
|
||||
if ($isRelevant && $score < 0.65) {
|
||||
$score = 0.65;
|
||||
$level = 'partially_related';
|
||||
}
|
||||
|
||||
$score = $this->snapScore($score, $bands);
|
||||
$level = $this->levelFromScore($score, $level);
|
||||
|
||||
return [$score, $level, $isRelevant, $role];
|
||||
}
|
||||
|
||||
private function enforceCombinedConsistency($combinedScore, $combinedLevel, $combinedRelevant, array $bands)
|
||||
{
|
||||
$combinedScore = $this->snapScore(floatval($combinedScore), $bands);
|
||||
$combinedLevel = $this->levelFromScore($combinedScore, $combinedLevel);
|
||||
$combinedRelevant = $combinedScore >= 0.65 - 0.001;
|
||||
|
||||
return [$combinedScore, $combinedLevel, $combinedRelevant];
|
||||
}
|
||||
|
||||
private function syncCombinedFieldsAcrossGroup(array $out)
|
||||
{
|
||||
$groups = [];
|
||||
foreach ($out as $idx => $row) {
|
||||
$key = (string)$row['cite_group_refs'];
|
||||
if ($key === '') {
|
||||
$key = 'ref:' . $row['reference_no'];
|
||||
}
|
||||
$groups[$key][] = $idx;
|
||||
}
|
||||
|
||||
foreach ($groups as $indices) {
|
||||
if (count($indices) <= 1) {
|
||||
continue;
|
||||
}
|
||||
$bestIdx = $indices[0];
|
||||
$bestScore = floatval($out[$bestIdx]['combined_relevance_score']);
|
||||
foreach ($indices as $idx) {
|
||||
$s = floatval($out[$idx]['combined_relevance_score']);
|
||||
if ($s >= $bestScore) {
|
||||
$bestScore = $s;
|
||||
$bestIdx = $idx;
|
||||
}
|
||||
}
|
||||
$src = $out[$bestIdx];
|
||||
foreach ($indices as $idx) {
|
||||
$out[$idx]['combined_is_relevant'] = intval($src['combined_is_relevant']);
|
||||
$out[$idx]['combined_relevance_score'] = floatval($src['combined_relevance_score']);
|
||||
$out[$idx]['combined_relevance_level'] = (string)$src['combined_relevance_level'];
|
||||
$out[$idx]['combined_reason'] = (string)$src['combined_reason'];
|
||||
}
|
||||
}
|
||||
|
||||
return $out;
|
||||
}
|
||||
|
||||
private function getScoreBands()
|
||||
{
|
||||
return [0.15, 0.25, 0.45, 0.65, 0.78, 0.85, 0.92, 0.98];
|
||||
}
|
||||
|
||||
private function snapScore($score, array $bands)
|
||||
{
|
||||
foreach ($bands as $band) {
|
||||
if (abs($score - $band) < 0.001) {
|
||||
return $band;
|
||||
}
|
||||
}
|
||||
$nearest = $bands[0];
|
||||
$minDiff = abs($score - $nearest);
|
||||
foreach ($bands as $band) {
|
||||
$diff = abs($score - $band);
|
||||
if ($diff < $minDiff) {
|
||||
$minDiff = $diff;
|
||||
$nearest = $band;
|
||||
}
|
||||
}
|
||||
|
||||
return $nearest;
|
||||
}
|
||||
|
||||
private function levelFromScore($score, $levelHint = '')
|
||||
{
|
||||
$levelHint = strtolower(trim((string)$levelHint));
|
||||
$allowed = ['highly_related', 'partially_related', 'weakly_related', 'unrelated'];
|
||||
if (in_array($levelHint, $allowed, true)) {
|
||||
return $levelHint;
|
||||
}
|
||||
$aliases = [
|
||||
'highly_related' => ['highly_related', 'high_related', 'strong_related', 'strong_relevance'],
|
||||
'partially_related' => ['partially_related', 'partial_related', 'moderate_related'],
|
||||
'weakly_related' => ['weakly_related', 'weak_related', 'low_related', 'insufficient'],
|
||||
'unrelated' => ['unrelated', 'not_related', 'irrelevant', 'no_meaningful_relevance'],
|
||||
];
|
||||
foreach ($aliases as $canonical => $list) {
|
||||
if (in_array($levelHint, $list, true)) {
|
||||
return $canonical;
|
||||
}
|
||||
}
|
||||
$score = floatval($score);
|
||||
if ($score >= 0.85) {
|
||||
return 'highly_related';
|
||||
}
|
||||
if ($score >= 0.65) {
|
||||
return 'partially_related';
|
||||
}
|
||||
if ($score >= 0.45) {
|
||||
return 'weakly_related';
|
||||
}
|
||||
|
||||
return 'unrelated';
|
||||
}
|
||||
|
||||
private function normalizeRelevanceRole($role)
|
||||
{
|
||||
$role = strtolower(trim((string)$role));
|
||||
$map = [
|
||||
'primary_relevance' => ['primary_relevance', 'primary_support', 'primary'],
|
||||
'supplementary_relevance' => ['supplementary_relevance', 'supplementary_support', 'supplementary'],
|
||||
'minimal_relevance' => ['minimal_relevance', 'minimal_support', 'minimal'],
|
||||
'no_meaningful_relevance' => ['no_meaningful_relevance', 'no_meaningful_support', 'none'],
|
||||
];
|
||||
foreach ($map as $canonical => $aliases) {
|
||||
if ($role === $canonical || in_array($role, $aliases, true)) {
|
||||
return $canonical;
|
||||
}
|
||||
}
|
||||
|
||||
return 'no_meaningful_relevance';
|
||||
}
|
||||
|
||||
private function cleanReason($reason)
|
||||
{
|
||||
$reason = trim(preg_replace('/\s+/u', ' ', (string)$reason));
|
||||
return mb_substr($reason, 0, 2000);
|
||||
}
|
||||
|
||||
private function boolVal($v)
|
||||
{
|
||||
if (is_bool($v)) {
|
||||
return $v;
|
||||
}
|
||||
if (is_numeric($v)) {
|
||||
return intval($v) !== 0;
|
||||
}
|
||||
$s = strtolower(trim((string)$v));
|
||||
return in_array($s, ['1', 'true', 'yes', 'y'], true);
|
||||
}
|
||||
|
||||
private function postChat(array $payload)
|
||||
{
|
||||
$this->lastPostError = '';
|
||||
try {
|
||||
$ch = curl_init();
|
||||
curl_setopt($ch, CURLOPT_URL, $this->url);
|
||||
curl_setopt($ch, CURLOPT_POST, true);
|
||||
curl_setopt($ch, CURLOPT_POSTFIELDS, json_encode($payload, JSON_UNESCAPED_UNICODE));
|
||||
curl_setopt($ch, CURLOPT_RETURNTRANSFER, true);
|
||||
curl_setopt($ch, CURLOPT_CONNECTTIMEOUT, min(15, $this->timeout));
|
||||
curl_setopt($ch, CURLOPT_TIMEOUT, $this->timeout);
|
||||
$headers = ['Content-Type: application/json'];
|
||||
if ($this->apiKey !== '') {
|
||||
$headers[] = 'Authorization: Bearer ' . $this->apiKey;
|
||||
}
|
||||
curl_setopt($ch, CURLOPT_HTTPHEADER, $headers);
|
||||
$raw = curl_exec($ch);
|
||||
if ($raw === false) {
|
||||
$this->lastPostError = 'LLM curl error: ' . curl_error($ch);
|
||||
\think\Log::warning('ReferenceRelevanceLlm: ' . $this->lastPostError);
|
||||
curl_close($ch);
|
||||
return null;
|
||||
}
|
||||
$httpCode = intval(curl_getinfo($ch, CURLINFO_HTTP_CODE));
|
||||
curl_close($ch);
|
||||
if ($httpCode < 200 || $httpCode >= 300) {
|
||||
$snippet = mb_substr(trim((string)$raw), 0, 200);
|
||||
$this->lastPostError = 'LLM HTTP ' . $httpCode . ($snippet !== '' ? ': ' . $snippet : '');
|
||||
\think\Log::warning('ReferenceRelevanceLlm: ' . $this->lastPostError);
|
||||
return null;
|
||||
}
|
||||
$data = json_decode($raw, true);
|
||||
if (!is_array($data)) {
|
||||
$this->lastPostError = 'LLM response is not valid JSON';
|
||||
return null;
|
||||
}
|
||||
if (isset($data['choices'][0]['message']['content'])) {
|
||||
return (string)$data['choices'][0]['message']['content'];
|
||||
}
|
||||
if (isset($data['content'])) {
|
||||
return (string)$data['content'];
|
||||
}
|
||||
$this->lastPostError = 'LLM response missing content field';
|
||||
} catch (\Exception $e) {
|
||||
$this->lastPostError = 'LLM exception: ' . $e->getMessage();
|
||||
\think\Log::warning('ReferenceRelevanceLlm: ' . $this->lastPostError);
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
|
||||
private function parseJson($raw)
|
||||
{
|
||||
$raw = trim((string)$raw);
|
||||
if ($raw === '') {
|
||||
return null;
|
||||
}
|
||||
$raw = preg_replace('/^```[a-zA-Z]*\s*|```$/m', '', $raw);
|
||||
$raw = trim($raw);
|
||||
$decoded = json_decode($raw, true);
|
||||
if (is_array($decoded)) {
|
||||
return $decoded;
|
||||
}
|
||||
if (preg_match('/\{[\s\S]*\}/', $raw, $m)) {
|
||||
$decoded = json_decode($m[0], true);
|
||||
if (is_array($decoded)) {
|
||||
return $decoded;
|
||||
}
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user