feat(data_process): 问答对数据评测体系与质量分雷达图

- 三层评测：规则层沿用原五维规则分，语义层用本地 BGE 向量算问答/来源
  相关性，评审层复用生成模型按 rubric 打分（忠实度/正确性/清晰度等，
  区分 standard/reasoning/dpo 输出类型），任一层失败自动降级
- 组合分 = 规则 35% + 语义 20% + 评审 45%，缺层自动重归一
- 新增 results/evaluate-batch 批量评测接口，镜像批量重生成的并发、
  乐观锁与部分成功语义；生成阶段不再展示质量分
- 详情页与结果编辑页新增"数据评测"按钮和批量进度；质量分列悬停弹出
  雷达图浮窗（评审 5 维 + 语义 2 维、三层分项、评审理由）
- 手动编辑/恢复后重算规则与语义层并丢弃过期评审分，雷达图不再展示
  失效数据

backend/app/modules/data_process/algorithms/quality.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import hashlib
 import json
+import math
 import re
 import unicodedata
 from collections import Counter
@@ -341,3 +342,87 @@ def score_quality(
         flags=tuple(flags),
         fingerprint=fingerprint,
     )
+
+
+def _cosine_similarity(left: Sequence[float], right: Sequence[float]) -> float:
+    if not left or not right or len(left) != len(right):
+        return 0.0
+    dot = math.fsum(a * b for a, b in zip(left, right))
+    norm_left = math.sqrt(math.fsum(a * a for a in left))
+    norm_right = math.sqrt(math.fsum(b * b for b in right))
+    if not norm_left or not norm_right:
+        return 0.0
+    return dot / (norm_left * norm_right)
+
+
+def semantic_quality_scores(
+    record: Mapping[str, Any],
+    *,
+    source_content: str = "",
+    embed_model: Any = None,
+) -> dict[str, Any] | None:
+    """用本地嵌入向量计算语义相关性（0-100）。
+
+    返回 ``question_answer``（问题↔答案）、``answer_source``（答案↔来源，
+    无来源时缺省）与 ``overall``；嵌入模型不可用时返回 None 降级，不阻断流程。
+    """
+
+    try:
+        if embed_model is None:
+            from .embedding import semantic_embedding_model
+
+            embed_model = semantic_embedding_model()
+        if embed_model is None:
+            return None
+
+        question = normalize_text(
+            " ".join(
+                str(record.get(field) or "")
+                for field in ("instruction", "input")
+            )
+        )
+        answer = normalize_text(
+            str(record.get("output") or "") or str(record.get("chosen") or "")
+        )
+        source = normalize_text(source_content)
+        texts = [text for text in {question, answer, source} if text]
+        if not texts:
+            return None
+        vectors = {text: embed_model.get_text_embedding(text) for text in texts}
+    except Exception:
+        return None
+
+    scores: dict[str, Any] = {}
+    if question and answer:
+        scores["question_answer"] = round(
+            100 * max(0.0, _cosine_similarity(vectors[question], vectors[answer])), 2
+        )
+    if answer and source:
+        scores["answer_source"] = round(
+            100 * max(0.0, _cosine_similarity(vectors[answer], vectors[source])), 2
+        )
+    if not scores:
+        return None
+    scores["overall"] = round(sum(scores.values()) / len(scores), 2)
+    return scores
+
+
+def composite_overall(
+    *,
+    rule: float | None,
+    semantic: float | None = None,
+    judge: float | None = None,
+) -> float:
+    """三层加权组合：规则 35% + 语义 20% + 评审 45%，缺失层自动重归一。"""
+
+    if rule is None:
+        rule = 0.0
+    if judge is not None and semantic is not None:
+        overall = rule * 0.35 + semantic * 0.20 + judge * 0.45
+    elif semantic is not None:
+        overall = rule * 0.60 + semantic * 0.40
+    elif judge is not None:
+        overall = rule * 0.55 + judge * 0.45
+    else:
+        overall = rule
+    return round(max(0.0, min(100.0, overall)), 2)