Python
# -*- coding: utf-8 -*-
"""
Chroma keyword_search 端到端测试。
注意:本地嵌入式 Chroma(PersistentClient)不支持原生 BM25/FTS 检索 ——
新的 collection.search() Search 表达式引擎只在 Chroma Server / Cloud(Rust
分布式后端)实现,本地调用会抛 NotImplementedError("Search is not implemented
for Local Chroma")。旧的 $contains 子串匹配命中只能给 binary 1.0 分,被 reader
端 normalize_bm25 sigmoid 压成 ~0.039 后对融合几乎无贡献还稀释强语义结果,
因此 keyword_search 在 chroma 后端故意 return [](诚实退化为纯向量召回)。
→ 这些断言"有命中"的用例对本地 chroma 不成立,整体 skip。真关键词检索请用
支持 sparse/BM25 的后端(qdrant)。chromadb>=1.5.9 已具备 server
模式的 BM25 能力,将来上 chroma server 时可重新启用。
"""
import asyncio
import random
import uuid
from typing import List
import pytest
from hy_memory.config import MemoryConfig
from hy_memory.models.memory import MemoryNode, MemoryLayer, MemoryStatus
pytestmark = pytest.mark.skip(
reason="本地 Chroma 不支持原生 BM25/FTS(Search 引擎仅 server/cloud 实现),"
"keyword_search 故意 return []。真关键词检索走 qdrant 后端。"
)
def run(coro):
loop = asyncio.new_event_loop()
try:
return loop.run_until_complete(coro)
finally:
loop.close()
def _make_node(
*,
content: str,
user_id: str = "u1",
agent_id: str = "a1",
session_id: str = "s1",
layer: MemoryLayer = MemoryLayer.L2_FACT,
status: MemoryStatus = MemoryStatus.ACTIVE,
is_latest: bool = True,
dims: int = 32,
) -> MemoryNode:
node = MemoryNode(
node_id=str(uuid.uuid4()),
user_id=user_id,
agent_id=agent_id,
session_id=session_id,
content=content,
layer=layer,
status=status,
is_latest=is_latest,
)
vec = [random.gauss(0, 1) for _ in range(dims)]
norm = sum(v * v for v in vec) ** 0.5
node.embedding = [v / norm for v in vec]
return node
@pytest.fixture
def store(tmp_path):
config = MemoryConfig.from_dict({
"vector_store": {
"provider": "chroma",
"collection_name": f"test_kw_{uuid.uuid4().hex[:8]}",
"persist_directory": str(tmp_path / "chroma_data"),
"embedding_dims": 32,
},
})
from hy_memory.data.vector_store_chroma import ChromaVectorStore
vs = ChromaVectorStore(config)
run(vs.initialize())
yield vs
run(vs.close())
class TestChromaKeywordSearchChinese:
def test_chinese_query_hits_chinese_content(self, store):
nodes = [
_make_node(content="用户倾向于接受务实的解决方案,主动寻找更新", user_id="u1"),
_make_node(content="今天天气不错,公园里很多人散步", user_id="u1"),
_make_node(content="用户喜欢吃辣,川菜火锅都行", user_id="u1"),
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(query="用户倾向接受务实", user_id="u1", top_k=5))
ids = {h["node_id"] for h in hits}
# 命中第 1 条(含"用户/倾向/接受/务实"),可能也命中第 3 条(含"用户")
assert nodes[0].node_id in ids
# 第 2 条不应命中(没"用户/倾向/接受/务实"任一)
assert nodes[1].node_id not in ids
def test_chinese_no_overlap_returns_empty(self, store):
nodes = [
_make_node(content="今天天气不错,公园里很多人", user_id="u1"),
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(query="完全无关的另一个查询", user_id="u1", top_k=5))
assert hits == []
def test_score_is_binary_one(self, store):
"""跟 Qdrant 行为对齐:所有命中 score=1.0"""
nodes = [_make_node(content="用户倾向接受务实", user_id="u1")]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(query="用户", user_id="u1", top_k=5))
assert len(hits) == 1
assert hits[0]["score"] == 1.0
class TestChromaKeywordSearchEnglish:
def test_english_query_hits(self, store):
nodes = [
_make_node(content="How to deploy to production with twine upload", user_id="u1"),
_make_node(content="User likes coffee in the morning", user_id="u1"),
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(query="deploy production", user_id="u1", top_k=5))
ids = {h["node_id"] for h in hits}
assert nodes[0].node_id in ids
assert nodes[1].node_id not in ids
class TestChromaKeywordSearchFilters:
def test_user_id_filter(self, store):
nodes = [
_make_node(content="用户倾向接受务实", user_id="u1"),
_make_node(content="用户倾向接受务实", user_id="u2"),
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(query="用户倾向", user_id="u1", top_k=10))
ids = {h["node_id"] for h in hits}
assert nodes[0].node_id in ids
assert nodes[1].node_id not in ids
def test_agent_id_filter(self, store):
nodes = [
_make_node(content="用户倾向接受务实", user_id="u1", agent_id="a1"),
_make_node(content="用户倾向接受务实", user_id="u1", agent_id="a2"),
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(
query="用户倾向", user_id="u1", agent_ids=["a1"], top_k=10,
))
ids = {h["node_id"] for h in hits}
assert nodes[0].node_id in ids
assert nodes[1].node_id not in ids
def test_layer_filter(self, store):
nodes = [
_make_node(content="用户倾向接受务实", user_id="u1", layer=MemoryLayer.L2_FACT),
_make_node(content="用户倾向接受务实", user_id="u1", layer=MemoryLayer.L4_IDENTITY),
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(
query="用户倾向", user_id="u1", layers=[MemoryLayer.L4_IDENTITY], top_k=10,
))
ids = {h["node_id"] for h in hits}
assert nodes[1].node_id in ids
assert nodes[0].node_id not in ids
def test_status_filter_excludes_non_active(self, store):
nodes = [
_make_node(content="用户倾向接受务实", user_id="u1", status=MemoryStatus.ACTIVE),
_make_node(content="用户倾向接受务实", user_id="u1", status=MemoryStatus.SUPERSEDED),
]
run(store.upsert_batch(nodes))
# 默认只搜 active
hits = run(store.keyword_search(query="用户倾向", user_id="u1", top_k=10))
ids = {h["node_id"] for h in hits}
assert nodes[0].node_id in ids
assert nodes[1].node_id not in ids
def test_only_latest(self, store):
nodes = [
_make_node(content="用户倾向接受务实", user_id="u1", is_latest=True),
_make_node(content="用户倾向接受务实", user_id="u1", is_latest=False),
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(
query="用户倾向", user_id="u1", only_latest=True, top_k=10,
))
ids = {h["node_id"] for h in hits}
assert nodes[0].node_id in ids
assert nodes[1].node_id not in ids
class TestChromaKeywordSearchEdgeCases:
def test_empty_query_returns_empty(self, store):
hits = run(store.keyword_search(query="", user_id="u1"))
assert hits == []
def test_whitespace_only_query_returns_empty(self, store):
hits = run(store.keyword_search(query=" ", user_id="u1"))
assert hits == []
def test_top_k_caps_results(self, store):
nodes = [
_make_node(content=f"用户倾向 候选 {i}", user_id="u1") for i in range(5)
]
run(store.upsert_batch(nodes))
hits = run(store.keyword_search(query="用户倾向", user_id="u1", top_k=3))
assert len(hits) <= 3