一处原文引用20多个参考文献优化

application/common/service/ReferenceRelevanceLlmService.php

@@ -19,6 +19,8 @@ class ReferenceRelevanceLlmService
     private $maxReferChars;
     private $maxAbstractChars;
     private $maxTokens;
+    /** 本次校对的联合引用组总篇数（分块时仍为整组篇数，供单篇分摊评分使用） */
+    private $groupRefCount = 0;
 
     public function __construct()
     {
@@ -72,6 +74,7 @@ class ReferenceRelevanceLlmService
         }
 
         $refCount = $this->countCiteGroupRefs($citeGroupRefs);
+        $this->groupRefCount = $refCount;
         // 默认每批最多 4 篇，降低单次排队/超时风险（可用 env 覆盖）
         $maxRefsPerCall = max(2, intval(Env::get('promotion.relevance_llm_max_refs_per_call', 4)));
         if ($refCount > $maxRefsPerCall) {
@@ -111,7 +114,8 @@ class ReferenceRelevanceLlmService
         $citeGroupRefs,
         array $referTypeMap,
         $refCount,
-        array $fallback
+        array $fallback,
+        array $fixedClaims = []
     ) {
         $payload = [
             'model'       => $this->model,
@@ -119,7 +123,7 @@ class ReferenceRelevanceLlmService
             'max_tokens'  => $this->resolveMaxTokens($refCount),
             'messages'    => [
                 ['role' => 'system', 'content' => $this->buildSystemPrompt()],
-                ['role' => 'user', 'content' => $this->buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs, $refCount, $referTypeMap)],
+                ['role' => 'user', 'content' => $this->buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs, $refCount, $referTypeMap, $fixedClaims)],
             ],
         ];
 
@@ -258,7 +262,8 @@ class ReferenceRelevanceLlmService
    - 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. **联合分写在顶层 combined_relevance_score / combined_reason**，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）；**禁止在 results 各条中重复 combined_***。
-   **联合分不得与单条分矛盾**：若全部单条均为 0.25，联合必须为 0.25；若全部单条均 ≤0.45，联合不得 >0.45；联合分不得远高于最高单条分（禁止「单条全弱、联合高分」）。
+   **联合分不得与单条分矛盾**：若全部单条均为 0.25，联合必须为 0.25；若全部单条均 ≤0.45，联合不得 >0.45（此处「单条弱」指主语不一致/证据层级不足等**实质弱相关**，不含规则18的分工覆盖情形）。
+   **但联合分不是「取单篇最高分」**：各篇分工覆盖不同 Claim 时，联合按规则18的整组覆盖并集定档，可高于任一单篇。
 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
@@ -301,7 +306,14 @@ class ReferenceRelevanceLlmService
     - 仅间接沾边、层级或主语明显不符 → **≤0.65**
     - **联合分**：多篇经典教材已系统覆盖全部并列功能时，联合可 **0.98**，不得因个别子类/哲学文献偏窄而压低整体
 14. **分值必须与你写出的 Claim 覆盖自洽（写完 reason 后必须自检；通用规则，适用所有文献类型）**：
-    - 先按你写出的覆盖标注计算「覆盖比例」= (完整✔数 + 0.5×部分数) / Claim总数，再据此定分：
+    - **先判本组是单独引用还是联合引用**：cite_group_refs 含多篇时按下方「联合引用分摊原则」评分；仅 1 篇时才按整段 Claim 总数算覆盖比例。
+    - **联合引用分摊原则（多篇联合时优先适用）**：整段 Claim 由组内各篇**分工覆盖**，**不要求任何单篇覆盖全部 Claim**。
+      - 单篇评分只看「**它覆盖的那几项 Claim 支撑得是否充分**」，**不看**「占整段 Claim 的比例」
+      - **主语一致 + 完整✔至少 1 项 Claim** → 单篇 **不低于 0.85**（已充分承担分工），**严禁**因未覆盖他篇负责的 Claim 而降到 0.45/0.25
+      - 主语一致但只有「部分」、无任何完整✔ → **0.65**
+      - 单篇**上限 0.85**：只承担整段一两项分工者不给 0.92/0.98（0.92+ 留给单篇即覆盖整段绝大多数 Claim 的情形）
+      - 组内多篇覆盖同一项 Claim 属**重复引用**，各篇分值照常给，但须在 combined_reason 点名
+    - 单独引用（组内仅 1 篇）时，按你写出的覆盖标注计算「覆盖比例」= (完整✔数 + 0.5×部分数) / Claim总数，再据此定分：
       - 覆盖比例 ≥0.9 且无✘ → **0.92~0.98**（系统/全部覆盖=0.98）
       - ≥0.72 → **0.92**；≥0.58 → **0.85**；≥0.45 → **0.78**；≥0.28 → **0.65**；≥0.15 → **0.45**；几乎全✘/无覆盖 → 0.25/0.15
     - **0.25/0.15 仅用于「主语不一致」或「几乎完全无覆盖（无任何✔/部分）」**。只要**主语一致**且 reason 写出**任意 ✔ 或部分**，就**禁止**给 0.25，至少 **0.45**。
@@ -327,6 +339,14 @@ class ReferenceRelevanceLlmService
     - **典型示例（多事实背景句 + 单篇原始研究）**：引用处「HF诊疗取得进展 + 老年HF患者QOL较低 + 低于其他慢病老人 + 因长疗程/并发症/急性加重」，文献为 Lee & Song 2015 SEM 症状管理与QOL研究：
       - Claim 应拆为：A诊疗进展✘、B HF患者QOL相关✔或部分、C 低于其他慢病比较✘、D 具体成因列举部分或✘
       - 覆盖约 35%~45% → **0.65（partially_related）**；**不得**因「都研究HF+QOL」给 0.92，**不得**因有多项✘就给 0.25
+18. **联合分按「整组 Claim 覆盖并集」定档（联合引用必用）**：
+    - 把组内各篇的 ✔/部分**取并集**：某项 Claim 只要有任一篇 ✔ 即记完整，只有「部分」记 0.5，无任何篇覆盖记 ✘
+    - 整组覆盖率 = (并集完整数 + 0.5×并集部分数) / Claim 总数，据此定档：
+      全部 Claim 均被覆盖 → **0.92~0.98**；≥0.72 → **0.85**；≥0.58 → **0.78**；≥0.28 → **0.65**；更低按实际
+    - **联合分与单篇分母不同**：单篇按分工评（规则14），联合按整段覆盖评；单篇偏低不构成压低联合分的理由
+    - **combined_reason 必须点名两件事**（这是本次校对对作者最有价值的输出）：
+      ①**无人覆盖的 Claim** → 写「X 暂无文献支撑」，提示作者该处需替换/改引
+      ②**被多篇重复覆盖的 Claim** → 写「X 由文献 a、b、c 重复覆盖」，提示引用堆砌可精简
 
 ==================================================
 【一、必须先拆解 claim】
@@ -489,7 +509,7 @@ Claim 具体内容见顶层 `claims`，combined_reason **只用字母指代**（
 PROMPT;
     }
 
-    private function buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs, $refCount = 0, array $referTypeMap = [])
+    private function buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs, $refCount = 0, array $referTypeMap = [], array $fixedClaims = [])
     {
         $parts = ["【正文节 t_article_main】\n" . $sectionText];
         if (trim((string)$citeGroupRefs) !== '') {
@@ -499,6 +519,14 @@ PROMPT;
         if ($localContext !== '') {
             $parts[] = "【本引用位置附近上下文（优先据此拆解 claim）】\n" . $localContext;
         }
+        if (!empty($fixedClaims)) {
+            $claimLines = [];
+            foreach ($fixedClaims as $letter => $text) {
+                $claimLines[] = $letter . '：' . $text;
+            }
+            $parts[] = "【已确定的 Claim 列表（本引用位置已拆解完成，必须沿用相同字母与含义，不得重新编号、改写或增删）】\n"
+                . implode("\n", $claimLines);
+        }
         $typeBlock = $this->formatReferTypeBlock($referTypeMap);
         if ($typeBlock !== '') {
             $parts[] = $typeBlock;
@@ -534,6 +562,9 @@ PROMPT;
         } else {
             $tail .= ' 单条引用：reason ≤220 字，点名各 Claim 覆盖与分值依据；顶层 combined_* 与单条一致。';
         }
+        if ($refCount > 1) {
+            $tail .= ' 联合引用分摊原则（务必遵守）：整段 Claim 由各篇分工覆盖，单篇只要主语一致且完整✔至少 1 项 Claim 即不低于 0.85（上限 0.85），严禁因未覆盖他篇负责的 Claim 而降到 0.45/0.25；仅「部分」覆盖则 0.65。顶层 combined_relevance_score 按各篇覆盖并集算整组覆盖率定档（全覆盖 0.92~0.98、≥0.72 给 0.85、≥0.58 给 0.78、≥0.28 给 0.65），combined_reason 必须点名「哪些 Claim 无任何文献支撑」与「哪些 Claim 被多篇重复覆盖（写明文献编号）」。';
+        }
         $parts[] = $tail;
 
         return implode("\n\n", $parts);
@@ -732,7 +763,8 @@ PROMPT;
                 $chunkRefs,
                 $chunkTypeMap,
                 count($chunk),
-                $fallback
+                $fallback,
+                $claims
             );
             if (!empty($part['request_failed']) || empty($part['results'])) {
                 $reason = isset($part['reason']) ? (string)$part['reason'] : 'LLM split batch failed';
@@ -771,6 +803,12 @@ PROMPT;
         if (count($results) > 1) {
             $bands = $this->getScoreBands();
             $adjustedCombined = $this->enforceCombinedAgainstSingles($results, $combinedScore, $bands);
+            // 各批只看到局部，联合分与说明按全组覆盖并集重算
+            $coverage = $this->summarizeGroupCoverage($results);
+            if ($this->maxSingleRelevanceScore($results) >= 0.65 - 0.001
+                && $coverage['floor'] > $adjustedCombined) {
+                $adjustedCombined = $coverage['floor'];
+            }
             if ($combinedScore <= 0 || abs($adjustedCombined - $combinedScore) > 0.001) {
                 $combinedScore = $adjustedCombined;
                 $combinedReason = $this->fallbackReasonFromScore(
@@ -778,6 +816,7 @@ PROMPT;
                     $this->levelFromScore($combinedScore)
                 );
             }
+            $combinedReason = $this->appendGroupCoverageNote($combinedReason, $coverage);
         } elseif (count($results) === 1) {
             $combinedScore = floatval($results[0]['relevance_score'] ?? 0);
             $combinedReason = (string)($results[0]['reason'] ?? '');
@@ -939,6 +978,18 @@ PROMPT;
                 $refLit = $this->extractRefLiteratureFromCombined($abstractText, $refNo);
                 $adjusted = $this->reconcileScoreAgainstCoverage($score, $reason, $refLit);
                 $capped = $this->reconcileScoreCeiling($adjusted, $reason);
+                // 联合引用分摊保底：整段 Claim 由多篇分工覆盖时，按整段分母算出的低分与✘数封顶均不适用
+                $shareFloor = $this->resolveJointShareFloor($reason);
+                if ($shareFloor > $capped) {
+                    \think\Log::warning(sprintf(
+                        'ReferenceRelevanceLlm: ref#%d joint share floor %.2f->%.2f (group=%d)',
+                        $refNo,
+                        $capped,
+                        $shareFloor,
+                        $this->groupRefCount
+                    ));
+                    $capped = $shareFloor;
+                }
                 if ($capped > $score + 0.001 || $capped < $score - 0.001) {
                     if ($capped > $score + 0.001) {
                         \think\Log::warning(sprintf(
@@ -1070,8 +1121,16 @@ PROMPT;
             }
         }
 
+        // 联合分按整组覆盖并集定档：各篇分工覆盖时不受单篇分母稀释拖累
+        $coverage = $this->summarizeGroupCoverage($outRows);
+        if ($this->maxSingleRelevanceScore($outRows) >= 0.65 - 0.001
+            && $coverage['floor'] > $combinedScore) {
+            $combinedScore = $coverage['floor'];
+        }
+
         list($combinedScore,) = $this->enforceCombinedConsistency($combinedScore, '', $bands);
         $combinedReason = $this->reconcileReasonScore($this->cleanReason($combinedReason), $combinedScore);
+        $combinedReason = $this->appendGroupCoverageNote($combinedReason, $coverage);
 
         return [
             'combined_relevance_score' => $combinedScore,
@@ -1372,6 +1431,217 @@ PROMPT;
         ];
     }
 
+    /**
+     * 解析 reason 中的 Claim 覆盖标注，返回 字母 => full|partial|fail。
+     * 同一字母多次出现时取最强标注（完整 > 部分 > 未覆盖）。
+     */
+    private function extractCoverageMarks($reason)
+    {
+        $reason = (string)$reason;
+        $covText = $reason;
+        if (preg_match('/Claim\s*覆盖[：:]\s*(.+?)(?:。|$)/us', $reason, $m)) {
+            $covText = $m[1];
+        }
+
+        $marks = [];
+        if (preg_match_all('/([A-Z])(?:\s*(✔|✘|×|部分))/u', $covText, $tokens, PREG_SET_ORDER)) {
+            foreach ($tokens as $tok) {
+                $letter = $tok[1];
+                $mark = $tok[2] === '✔' ? 'full' : ($tok[2] === '部分' ? 'partial' : 'fail');
+                $cur = isset($marks[$letter]) ? $marks[$letter] : '';
+                if ($cur === 'full' || ($cur === 'partial' && $mark === 'fail')) {
+                    continue;
+                }
+                $marks[$letter] = $mark;
+            }
+        }
+
+        return $marks;
+    }
+
+    /**
+     * 联合引用中，整段 Claim 由各篇分工覆盖，单篇不应因未覆盖他篇负责的 Claim 而被判弱相关。
+     * 按「分摊份额」（本篇覆盖数 ÷ 人均应覆盖数）给一个保底分。
+     *
+     * @return float 保底分；0 表示不适用（单独引用、主语不一致、无覆盖标注等）
+     */
+    private function resolveJointShareFloor($reason)
+    {
+        $groupSize = intval($this->groupRefCount);
+        if ($groupSize < 2) {
+            return 0.0;
+        }
+        $reason = (string)$reason;
+        if ($reason === '' || !preg_match('/主语一致/u', $reason)) {
+            return 0.0;
+        }
+        // 主语/层级/类型不适配属实质弱相关，分摊原则不适用
+        if (preg_match('/主语不一致|主语层级|层级不对|层级不一致|类型不符合|类型不适配|不适合支撑/ui', $reason)) {
+            return 0.0;
+        }
+
+        $cov = $this->countCoverageSignals($reason);
+        $full = $cov['full'];
+        $partial = $cov['partial'];
+        $total = $cov['total'];
+        if ($total < 1 || ($full === 0 && $partial === 0)) {
+            return 0.0;
+        }
+
+        $expected = $total / $groupSize;
+        if ($expected <= 0) {
+            return 0.0;
+        }
+        $shareRatio = min(1.0, ($full + 0.5 * $partial) / $expected);
+
+        if ($shareRatio >= 0.72) {
+            $floor = 0.92;
+        } elseif ($shareRatio >= 0.58) {
+            $floor = 0.85;
+        } elseif ($shareRatio >= 0.45) {
+            $floor = 0.78;
+        } elseif ($shareRatio >= 0.28) {
+            $floor = 0.65;
+        } else {
+            return 0.0;
+        }
+
+        // 完成分摊份额不等于覆盖全段：完整覆盖至少一项封顶 0.85，仅部分覆盖封顶 0.65
+        $ceiling = $full >= 1 ? 0.85 : 0.65;
+
+        return $this->snapScore(min($floor, $ceiling), $this->getScoreBands());
+    }
+
+    /**
+     * 汇总整组 Claim 覆盖情况：各篇覆盖取并集，得出整组覆盖率、无人覆盖项与重复覆盖项。
+     *
+     * @return array{total:int,full:int,partial:int,ratio:float,floor:float,missing:array,duplicated:array}
+     */
+    private function summarizeGroupCoverage(array $results)
+    {
+        $empty = [
+            'total'      => 0,
+            'full'       => 0,
+            'partial'    => 0,
+            'ratio'      => 0.0,
+            'floor'      => 0.0,
+            'missing'    => [],
+            'duplicated' => [],
+        ];
+        if (count($results) < 2) {
+            return $empty;
+        }
+
+        $union = [];
+        $coveredBy = [];
+        foreach ($results as $row) {
+            $reason = isset($row['reason']) ? (string)$row['reason'] : '';
+            if ($reason === '') {
+                continue;
+            }
+            $refNo = intval(isset($row['reference_no']) ? $row['reference_no'] : 0);
+            foreach ($this->extractCoverageMarks($reason) as $letter => $mark) {
+                $cur = isset($union[$letter]) ? $union[$letter] : 'fail';
+                if ($mark === 'full' || $cur === 'full') {
+                    $union[$letter] = 'full';
+                } elseif ($mark === 'partial' || $cur === 'partial') {
+                    $union[$letter] = 'partial';
+                } else {
+                    $union[$letter] = 'fail';
+                }
+                if ($mark === 'full' && $refNo > 0) {
+                    $coveredBy[$letter][] = $refNo;
+                }
+            }
+        }
+
+        $total = count($union);
+        if ($total < 1) {
+            return $empty;
+        }
+        ksort($union, SORT_STRING);
+
+        $full = $partial = 0;
+        $missing = [];
+        foreach ($union as $letter => $mark) {
+            if ($mark === 'full') {
+                $full++;
+            } elseif ($mark === 'partial') {
+                $partial++;
+            } else {
+                $missing[] = $letter;
+            }
+        }
+        if ($full === 0 && $partial === 0) {
+            return $empty;
+        }
+
+        $ratio = ($full + 0.5 * $partial) / $total;
+        if ($ratio >= 0.9 && $full === $total) {
+            $floor = 0.92;
+        } elseif ($ratio >= 0.72) {
+            $floor = 0.85;
+        } elseif ($ratio >= 0.58) {
+            $floor = 0.78;
+        } elseif ($ratio >= 0.28) {
+            $floor = 0.65;
+        } else {
+            $floor = 0.0;
+        }
+
+        $duplicated = [];
+        foreach ($coveredBy as $letter => $refNos) {
+            $refNos = array_values(array_unique($refNos));
+            if (count($refNos) > 1) {
+                sort($refNos, SORT_NUMERIC);
+                $duplicated[$letter] = $refNos;
+            }
+        }
+        ksort($duplicated, SORT_STRING);
+
+        return [
+            'total'      => $total,
+            'full'       => $full,
+            'partial'    => $partial,
+            'ratio'      => $ratio,
+            'floor'      => $floor > 0 ? $this->snapScore($floor, $this->getScoreBands()) : 0.0,
+            'missing'    => $missing,
+            'duplicated' => $duplicated,
+        ];
+    }
+
+    /**
+     * 在 combined_reason 末尾追加整组覆盖说明（缺口 Claim 与重复覆盖 Claim）。
+     * 分块校对时各批会各写一次，故先清除旧标记再按全组重写。
+     */
+    private function appendGroupCoverageNote($combinedReason, array $summary)
+    {
+        $combinedReason = trim((string)$combinedReason);
+        $combinedReason = trim(preg_replace('/【整组覆盖】.*$/us', '', $combinedReason));
+        if (intval($summary['total']) < 1) {
+            return $combinedReason;
+        }
+
+        $segments = [sprintf('Claim 覆盖 %d/%d', $summary['full'], $summary['total'])];
+        if (!empty($summary['missing'])) {
+            $segments[] = implode('、', $summary['missing']) . ' 暂无文献支撑，建议替换相应编号文献或调整该句表述';
+        }
+        if (!empty($summary['duplicated'])) {
+            $dup = [];
+            foreach ($summary['duplicated'] as $letter => $refNos) {
+                $dup[] = $letter . ' 由文献 ' . implode('、', $refNos) . ' 重复覆盖';
+            }
+            $segments[] = implode('；', $dup) . '，可酌情精简';
+        }
+        if (count($segments) < 2) {
+            return $combinedReason;
+        }
+
+        $note = '【整组覆盖】' . implode('；', $segments) . '。';
+
+        return $combinedReason === '' ? $note : $combinedReason . ' ' . $note;
+    }
+
     /**
      * 从联合文献块中提取指定编号的摘要/清洗内容。
      */