阶段3（下）：RAG、向量数据库专题；接入 BOSS直聘调研（research/、岗位对照和延伸技术栈）；新增 cs 领域和 12 个技术栈骨架；面试题全部补齐追问并加入校验

topics/ai/vector-db/examples/vector-stores/common.py

@@ -0,0 +1,33 @@
+"""示例共用：确定性的“假嵌入”（随机向量 + 主题偏移），让示例不依赖嵌入模型 API。"""
+import numpy as np
+
+DIM = 64
+TOPICS = ["hr", "it", "finance"]
+
+
+def make_corpus(n_per_topic: int = 200, seed: int = 7):
+    """每个主题一个中心向量，文档向量 = 中心 + 噪声。返回 (ids, vectors, metadatas)。"""
+    rng = np.random.default_rng(seed)
+    centers = {t: rng.normal(size=DIM) for t in TOPICS}
+    ids, vecs, metas = [], [], []
+    for t in TOPICS:
+        for i in range(n_per_topic):
+            v = centers[t] + rng.normal(scale=0.8, size=DIM)
+            ids.append(f"{t}-{i}")
+            vecs.append(v / np.linalg.norm(v))
+            metas.append({"topic": t, "tenant": f"tenant-{i % 3}"})
+    return ids, np.array(vecs, dtype=np.float32), metas, centers
+
+
+def query_vec(centers, topic: str, seed: int = 1) -> np.ndarray:
+    rng = np.random.default_rng(seed)
+    v = centers[topic] + rng.normal(scale=0.8, size=DIM)
+    return (v / np.linalg.norm(v)).astype(np.float32)
+
+
+def brute_force(vecs: np.ndarray, q: np.ndarray, k: int, mask=None) -> list[int]:
+    """精确的最近邻（余弦 = 归一化后的点积），作为近似检索的对照。"""
+    scores = vecs @ q
+    if mask is not None:
+        scores = np.where(mask, scores, -np.inf)
+    return list(np.argsort(-scores)[:k])