"""在已算好的候选得分上比较两种筛选顺序。"""
candidates = [
    {"source": "bob-private", "owner": "bob", "score": .99},
    {"source": "numpy-public", "owner": "public", "score": .95},
    {"source": "alice-plan", "owner": "alice", "score": .80},
]
# Alice 的候选集合先确定为公共材料和自己的材料。
allowed = [item for item in candidates if item["owner"] in ("public", "alice")]
correct = sorted(allowed, key=lambda item: item["score"], reverse=True)[:2]
# 若先在全库截断到两个结果，再删除无权读取的项，就少取了一份可用材料。
global_two = sorted(candidates, key=lambda item: item["score"], reverse=True)[:2]
late_filter = [item for item in global_two if item["owner"] in ("public", "alice")]
assert [item["source"] for item in correct] == ["numpy-public", "alice-plan"]
assert [item["source"] for item in late_filter] == ["numpy-public"]
print("先筛选：", correct, "先截断：", late_filter)

# 对照完整文件阅读这一段；此处保留当前步骤使用的函数或类型。
class NoteIndex:
    """小型语料重开时重新建立索引，业务计划与图检查点另行持久保存。"""
    def __init__(self, embeddings: Embeddings | None = None):
        self.chunks = load_chunks()
        self.store = InMemoryVectorStore.from_documents(self.chunks, embedding=embeddings or CharacterEmbeddings())

    def search(self, query: str, owner: str, k: int = 2) -> list[dict]:
        # 先按归属过滤，再选择 k 个结果，模型不能指定另一个用户。
        pairs = self.store.similarity_search_with_score(
            query, k=k, filter=lambda document: document.metadata["owner"] in ("public", owner))
        return [{"chunk_id": doc.metadata["chunk_id"], "source": doc.metadata["source"],
                 "text": doc.page_content, "score": float(score)} for doc, score in pairs if score >= 0.12]