update

ruoyi-fastapi-backend/module_admin/service/kb_es_service.py

@@ -1,5 +1,6 @@
 import hashlib
 import json
+import re
 import time
 from dataclasses import dataclass
 from typing import Any, Dict, List, Tuple
@@ -71,30 +72,18 @@ class KBESService:
         chunks, metrics_raw = await cls._es_search(question)
         retrieve_ms = (time.perf_counter() - t0) * 1000
 
-        match_service = get_match_service()
-        combined = "\n".join(
-            [
-                (c.get("question") or "")
-                + "\n"
-                + (c.get("answer") or "")
-                + "\n"
-                + (c.get("title") or "")
-                + "\n"
-                + (c.get("content") or "")
-                for c in chunks
-            ]
-        )
-        coverage = match_service.calculate_keyword_coverage(question, combined)
+        chunks, best_cov = cls._rerank_chunks(question, chunks)
 
-        top1 = metrics_raw.get("top1_score", 0.0)
-        gap = metrics_raw.get("score_gap", 0.0)
+        top1 = float(chunks[0].get("_score") or 0.0) if chunks else 0.0
+        top2 = float(chunks[1].get("_score") or 0.0) if len(chunks) > 1 else 0.0
+        gap = top1 - top2
 
         metrics = KBRetrievalMetrics(
             retrieve_ms=retrieve_ms,
             top1_score=top1,
             score_gap=gap,
             hit_count=len(chunks),
-            keyword_coverage=coverage,
+            keyword_coverage=best_cov,
             cache_hit=cache_hit,
         )
 
@@ -123,6 +112,63 @@ class KBESService:
 
         return chunks, metrics
 
+    @classmethod
+    def _rerank_chunks(cls, question: str, chunks: List[Dict[str, Any]]) -> Tuple[List[Dict[str, Any]], float]:
+        """二次重排（轻量、无LLM），避免 ES top1 误命中导致直接输出低质答案。
+
+        目标：尽量保持速度（topK 很小），只在内存里做简单特征打分。
+        """
+
+        if not chunks:
+            return chunks, 0.0
+
+        match_service = get_match_service()
+        q = (question or "").strip()
+        q_norm = q.lower()
+
+        # 约束词：如果用户问到公司/年份，命中片段最好也出现对应实体
+        must_brand = any(x in q_norm for x in ("康达新材", "康达"))
+        years = set(re.findall(r"20\d{2}", q))
+
+        max_es = max([float(c.get("_score") or 0.0) for c in chunks] or [0.0])
+        if max_es <= 0:
+            max_es = 1.0
+
+        scored: List[Tuple[float, float, float, Dict[str, Any]]] = []
+        for c in chunks:
+            text = "\n".join(
+                [
+                    str(c.get("question") or ""),
+                    str(c.get("answer") or ""),
+                    str(c.get("title") or ""),
+                    str(c.get("content") or ""),
+                    str(c.get("source") or ""),
+                ]
+            )
+            text_norm = text.lower()
+
+            cov = match_service.calculate_keyword_coverage(q, text)
+            es = float(c.get("_score") or 0.0)
+            es_norm = es / max_es
+
+            penalty = 0.0
+            if must_brand and ("康达" not in text_norm):
+                penalty -= 0.35
+            if years:
+                hit_years = any(y in text for y in years)
+                if not hit_years:
+                    penalty -= 0.25
+
+            # 覆盖率更能反映“是否问对文档”，ES 分数反映粗排相关性
+            final = 0.55 * cov + 0.45 * es_norm + penalty
+
+            scored.append((final, cov, es, c))
+
+        scored.sort(key=lambda x: x[0], reverse=True)
+        reranked = [c for _, _, _, c in scored]
+        best_cov = float(scored[0][1]) if scored else 0.0
+        return reranked, best_cov
+
     @classmethod
     def is_confident(cls, metrics: KBRetrievalMetrics) -> bool:
         if metrics.hit_count <= 0: