候选岗位补核心职责（赛题定义五要素齐备）；全岗位补2024-2026Q3近期演化信号并收敛未来窗口；MinHash升至256哈希修复边沿误杀；审核台展示职责/场景

app/cleaning/dedup.py

@@ -13,7 +13,7 @@ from app.models import JDRecord
 
 logger = logging.getLogger(__name__)
 
-NUM_HASHES = 128         # MinHash 哈希函数个数（>=64，越高估计越稳）
+NUM_HASHES = 256         # MinHash 哈希函数个数（越高估计越稳；128 时边沿样本偏差可达 ±0.1 误杀植入簇）
 SHINGLE_N = 5            # 字符 5-gram shingles
 LSH_BANDS = 16           # 16 bands × 8 rows（严格分带，抑制模板化文本误配）
 _PRIME = (1 << 61) - 1   # 梅森素数，用于 (a*x+b) mod p 哈希族

app/discover/cluster.py

@@ -23,6 +23,7 @@ MIN_COMPANIES = 5         # 独立企业数（防单一企业造词）
 MIN_GROWTH = 1.5          # 近期增长率（最近一年 JD 数 / 前一年，≥50% 增长）
 MIN_DISTINCTNESS = 0.4    # 与既有岗位差异度：1 - 与最近既有岗位质心的余弦相似度
 MERGE_SIM = 0.50          # 凝聚合并的质心余弦相似度阈值
+RESP_MIN_RATIO = 0.25     # 候选职责入选阈值：簇内 ≥25%（且 ≥2 条）JD 提及，取频次 Top 5
 
 
 # ---------- TF-IDF（字符 2/3-gram） ----------
@@ -158,21 +159,37 @@ def discover_new_jobs(
         if distinctness < min_distinctness:
             logger.info("簇与既有岗位差异度不足（%.2f < %.2f），不列为新岗位", distinctness, min_distinctness)
             continue
-        # 簇内最常见的抽取标题作为候选岗位名；技能取簇内高频
+        # 簇内最常见的抽取标题作为候选岗位名；技能/职责/场景取簇内高频
         titles = Counter(extractions[r.jid].job_title for r in group)
         title = titles.most_common(1)[0][0]
         req = Counter(s for r in group for s in extractions[r.jid].required_skills)
         pref = Counter(s for r in group for s in extractions[r.jid].preferred_skills)
+        duty = Counter(d for r in group for d in extractions[r.jid].responsibilities)
         scen = Counter(s for r in group for s in extractions[r.jid].scenarios)
         ind = Counter(r.industry for r in group if r.industry)
         n = len(group)
+        industries = [s for s, _ in ind.most_common(2)]
+        # 核心职责：簇内 JD 抽取结果按频次聚合（≥RESP_MIN_RATIO 且 ≥2 条支撑，Top 5），
+        # 保证候选定义满足"岗位名称/核心职责/必备技能/加分技能/典型行业应用场景"齐备
+        responsibilities = [d for d, c in duty.most_common(5) if c >= max(2, n * RESP_MIN_RATIO)]
+        required = [s for s, c in req.most_common() if c >= n * 0.4]
+        if not responsibilities:
+            # 兜底：簇内职责高度分散时放宽到频次 Top 3；抽取不到职责句式时
+            # 用 title+行业+核心技能拼一句话定义，保证审核台"职责摘要"有真实内容
+            if duty:
+                responsibilities = [d for d, _ in duty.most_common(3)]
+            else:
+                focus = "、".join(required[:2]) or title
+                responsibilities = [
+                    f"面向{'/'.join(industries) or '新兴'}行业，承担{title}相关的{focus}设计、开发与交付工作"]
         candidates.append(JobDefinition(
             job_id=f"cand_{title}",
             title=title,
             category="新岗位候选",
-            industries=[s for s, _ in ind.most_common(2)],
-            required_skills=[s for s, c in req.most_common() if c >= n * 0.4],
+            industries=industries,
+            required_skills=required,
             preferred_skills=[s for s, c in pref.most_common() if c >= n * 0.25],
+            responsibilities=responsibilities,
             scenarios=[s for s, _ in scen.most_common(5)],
             status="pending",
             source_count=n,