feat(knowledge): Executor 行动记忆 + 归一化质量等级解决 ISS-002 重复检索

- RetrievedDocTracker 升级为域级+文档级双层记录（Map<sessionId, Map<domain, Set<filePath>>>） - LookupKnowledgeTool 新增 Min-Max 归一化层（BGE-M3 L2 距离→[0,1] similarity） - 三等级 relevanceLevel：PRECISE / HIGHLY_RELEVANT / REFERENCE + completenessHint 兜底信号 - LookupResult 新增 relevanceLevel、completenessHint、retrievedDomainsThisSession - Executor prompt 重写：4 条检索约束 + 合法出口不查全不追责，重复检索才惩罚 - 入库可观测性：V010 迁移 + retrieval_details JSON 扩展 - 归档 executor-action-memory-relevance change
2026-07-01 18:24:41 +08:00
parent e4f37cb9e6
commit e438df4355
21 changed files with 1076 additions and 88 deletions
@@ -86,8 +86,15 @@ class FullPipelineSmokeTest {
        boolean hasNonZero = vector.stream().anyMatch(v -> Math.abs(v) > 1e-6);
        assertTrue(hasNonZero, "向量不能全为零");

+        // L2 范数校验：BGE-M3 输出应为 L2 归一化的单位向量
+        double norm = Math.sqrt(vector.stream().mapToDouble(v -> (double) v * v).sum());
+
        System.out.println("维度: " + vector.size());
        System.out.println("前5维: " + vector.subList(0, Math.min(5, vector.size())));
+        System.out.println("L2 范数: " + String.format("%.10f", norm));
+        System.out.println("是否归一化 (|norm - 1.0| < 0.01): " + (Math.abs(norm - 1.0) < 0.01));
+
+        assertEquals(1.0, norm, 0.01, "BGE-M3 向量应为 L2 归一化单位向量，实际范数=" + norm);
        System.out.println("Embedding ✓");
    }