329 lines
13 KiB
Python
329 lines
13 KiB
Python
# -*- coding: utf-8 -*-
|
|
"""
|
|
R2：调用方自定义 meta 的透传与 merge 追加。
|
|
|
|
需求口径（CTX）：
|
|
- meta 为 list，一个记忆节点可挂多份，每份代表「该次 write 时刻」的上下文
|
|
- add（新增记忆）：直接用本次传入的 meta；一次 write 产生多条新增时各条都带同一份
|
|
- merge（合并到已有记忆）：追加到旧节点的 meta list 上，**保留旧 meta，不覆盖**
|
|
- meta 不 embed、不参与召回、不作过滤条件
|
|
|
|
本文件锁死上述语义，以及「不传 meta 时与改动前逐字一致」这条向后兼容契约。
|
|
"""
|
|
|
|
import asyncio
|
|
import inspect
|
|
|
|
import pytest
|
|
|
|
from hy_memory.client import HyMemoryClient
|
|
from hy_memory.pipelines.base import WriteRequest
|
|
from hy_memory.pipelines.writer import MemoryWriter
|
|
|
|
|
|
def _run(coro):
|
|
loop = asyncio.new_event_loop()
|
|
try:
|
|
return loop.run_until_complete(coro)
|
|
finally:
|
|
loop.close()
|
|
|
|
|
|
def _req(meta=None):
|
|
return WriteRequest(content="x", client_meta=meta)
|
|
|
|
|
|
class _Op:
|
|
"""够用的 reconcile op 替身：_build_custom 只碰 supersede_reason。"""
|
|
|
|
def __init__(self, supersede_reason=None):
|
|
self.supersede_reason = supersede_reason
|
|
|
|
|
|
META1 = {"source_platform": "codebuddy", "has_tool_call": True, "scene": "chat"}
|
|
META2 = {"source_platform": "workbuddy", "has_tool_call": False, "scene": "review"}
|
|
|
|
|
|
class TestWriteRequestField:
|
|
def test_client_meta_defaults_to_none(self):
|
|
assert WriteRequest(content="x").client_meta is None
|
|
|
|
def test_client_meta_is_separate_from_extra(self):
|
|
"""extra 混放 SDK 控制键（agent_mode / mode / layer），合用会互相覆盖。"""
|
|
req = WriteRequest(content="x", extra={"mode": "pro"}, client_meta={"mode": "biz"})
|
|
assert req.extra == {"mode": "pro"}
|
|
assert req.client_meta == {"mode": "biz"}
|
|
|
|
|
|
class TestAddEntrypoint:
|
|
def test_async_add_exposes_meta(self):
|
|
params = inspect.signature(HyMemoryClient.async_add).parameters
|
|
assert "meta" in params
|
|
assert params["meta"].default is None
|
|
|
|
def test_sync_add_exposes_meta(self):
|
|
params = inspect.signature(HyMemoryClient.add).parameters
|
|
assert "meta" in params
|
|
assert params["meta"].default is None
|
|
|
|
|
|
class TestBuildCustomOnNewNode:
|
|
"""新建节点（ADD / SUPERSEDE / 直写 / summary / intention）只带本次 meta。"""
|
|
|
|
def test_meta_wrapped_in_list(self):
|
|
custom = MemoryWriter._build_custom(_Op(), "rid-1", _req(META1))
|
|
assert custom["meta"] == [META1], "单份 meta 也必须是 list，读侧才不用分支"
|
|
|
|
def test_no_meta_key_when_not_provided(self):
|
|
"""不传 meta 时 custom 必须与改动前逐字一致，不能多出空键。"""
|
|
custom = MemoryWriter._build_custom(_Op(), "rid-1", _req(None))
|
|
assert custom == {"request_id": "rid-1"}
|
|
|
|
def test_no_meta_key_when_request_absent(self):
|
|
assert MemoryWriter._build_custom(_Op(), "rid-1") == {"request_id": "rid-1"}
|
|
|
|
def test_empty_meta_treated_as_absent(self):
|
|
assert "meta" not in MemoryWriter._build_custom(_Op(), "rid-1", _req({}))
|
|
|
|
def test_coexists_with_internal_keys(self):
|
|
custom = MemoryWriter._build_custom(_Op("contradiction"), "rid-1", _req(META1))
|
|
assert custom == {
|
|
"request_id": "rid-1",
|
|
"supersede_reason": "contradiction",
|
|
"meta": [META1],
|
|
}
|
|
|
|
def test_meta_is_copied_not_referenced(self):
|
|
"""存进 custom 后再改原 dict 不应影响已落库内容。"""
|
|
src = dict(META1)
|
|
custom = MemoryWriter._build_custom(_Op(), "rid-1", _req(src))
|
|
src["scene"] = "mutated"
|
|
assert custom["meta"][0]["scene"] == "chat"
|
|
|
|
def test_op_may_be_none(self):
|
|
"""summary / intention 等路径没有 reconcile op。"""
|
|
custom = MemoryWriter._build_custom(None, "rid-1", _req(META1))
|
|
assert custom["meta"] == [META1]
|
|
|
|
|
|
class TestMergeAppendSemantics:
|
|
"""UPDATE 原地改同一节点，旧 meta 必须保留——这是 R2 的核心契约。"""
|
|
|
|
def test_appends_to_existing_chain(self):
|
|
merged = MemoryWriter._merge_client_meta({"meta": [META1]}, _req(META2))
|
|
assert merged["meta"] == [META1, META2], "旧 meta 必须保留在前，新的追加在后"
|
|
|
|
def test_accumulates_across_multiple_writes(self):
|
|
custom = {"meta": [META1]}
|
|
for i in range(3):
|
|
custom = MemoryWriter._merge_client_meta(custom, _req({"n": i}))
|
|
assert custom["meta"] == [META1, {"n": 0}, {"n": 1}, {"n": 2}]
|
|
|
|
def test_first_meta_on_node_without_any(self):
|
|
merged = MemoryWriter._merge_client_meta({"request_id": "old"}, _req(META1))
|
|
assert merged["meta"] == [META1]
|
|
|
|
def test_preserves_other_custom_keys(self):
|
|
"""custom 是整体替换写回的，内部字段不能被 meta 合并冲掉。"""
|
|
merged = MemoryWriter._merge_client_meta(
|
|
{"request_id": "old-rid", "basic_info_kv": {"name": "ivan"}, "meta": [META1]},
|
|
_req(META2),
|
|
)
|
|
assert merged["request_id"] == "old-rid"
|
|
assert merged["basic_info_kv"] == {"name": "ivan"}
|
|
assert merged["meta"] == [META1, META2]
|
|
|
|
def test_returns_none_when_no_meta_passed(self):
|
|
"""返回 None 是「无需回读与写回」的信号，保证未用该功能者零额外开销。"""
|
|
assert MemoryWriter._merge_client_meta({"meta": [META1]}, _req(None)) is None
|
|
|
|
def test_handles_missing_old_custom(self):
|
|
assert MemoryWriter._merge_client_meta(None, _req(META1))["meta"] == [META1]
|
|
|
|
def test_non_list_legacy_meta_is_not_dropped(self):
|
|
"""历史数据里 meta 可能是裸 dict；降级成首元素，不能静默丢。"""
|
|
merged = MemoryWriter._merge_client_meta({"meta": META1}, _req(META2))
|
|
assert merged["meta"] == [META1, META2]
|
|
|
|
def test_does_not_mutate_input(self):
|
|
old = {"meta": [META1]}
|
|
MemoryWriter._merge_client_meta(old, _req(META2))
|
|
assert old == {"meta": [META1]}, "回读到的旧 custom 不应被就地改写"
|
|
|
|
|
|
class TestUpdateAppendsThroughStore:
|
|
"""端到端走一遍存储：UPDATE 分支必须真的把追加后的 meta 写回节点。
|
|
|
|
上面的用例只覆盖纯函数，无法证明 writer 调用了它、也无法证明 custom 被写进
|
|
update_payload。R2 的核心风险恰恰在这里——漏写回就会静默丢掉旧 meta。
|
|
"""
|
|
|
|
DIMS = 32
|
|
|
|
def _emb(self):
|
|
import random
|
|
|
|
vec = [random.gauss(0, 1) for _ in range(self.DIMS)]
|
|
norm = sum(v * v for v in vec) ** 0.5 or 1.0
|
|
return [v / norm for v in vec]
|
|
|
|
def _store(self, tmp_path):
|
|
import uuid as _uuid
|
|
|
|
from hy_memory.config import MemoryConfig
|
|
from hy_memory.data.vector_store_faiss import FaissVectorStore
|
|
|
|
cfg = MemoryConfig.from_dict({
|
|
"vector_store": {
|
|
"provider": "faiss",
|
|
"collection_name": f"test_meta_{_uuid.uuid4().hex[:8]}",
|
|
"persist_directory": str(tmp_path / "faiss_meta"),
|
|
"embedding_dims": self.DIMS,
|
|
},
|
|
})
|
|
store = FaissVectorStore(cfg)
|
|
_run(store.initialize())
|
|
return cfg, store
|
|
|
|
def _writer(self, cfg, store, ops):
|
|
from hy_memory.agent.reconciler import ReconcileResult
|
|
from hy_memory.pipelines.writer import MemoryWriter
|
|
|
|
writer = MemoryWriter(cfg, vector_store=store)
|
|
_run(writer.initialize())
|
|
|
|
outer = self
|
|
|
|
class _FakeEmbed:
|
|
async def embed_batch(self, texts):
|
|
return [outer._emb() for _ in texts]
|
|
|
|
async def embed_queued(self, text, use_cache=True):
|
|
return outer._emb()
|
|
|
|
writer._embed_service = _FakeEmbed()
|
|
|
|
async def _fake_reconcile(*args, **kwargs):
|
|
return ReconcileResult(ops=ops)
|
|
|
|
writer.reconciler.reconcile = _fake_reconcile
|
|
return writer
|
|
|
|
def _seed(self, store, custom):
|
|
import uuid as _uuid
|
|
|
|
from hy_memory.models.memory import MemoryNode, MemoryLayer, MemoryStatus
|
|
|
|
node = MemoryNode(
|
|
node_id=str(_uuid.uuid4()), user_id="u1", agent_id="default_agent",
|
|
session_id="s1", content="用户用 React。", layer=MemoryLayer.L2_FACT,
|
|
status=MemoryStatus.ACTIVE, is_latest=True, owner="user",
|
|
custom=custom,
|
|
)
|
|
node.embedding = self._emb()
|
|
_run(store.upsert(node))
|
|
return node
|
|
|
|
def _update_op(self, node_id, content="用户用 React 18 + TS。"):
|
|
from hy_memory.agent.reconciler import ReconcileOp
|
|
|
|
return ReconcileOp(
|
|
op="UPDATE", memory_id=node_id, content=content,
|
|
owner="user", tags=[], layer="L2_FACT",
|
|
)
|
|
|
|
def _write(self, writer, store, ops, meta):
|
|
req = WriteRequest(content="x", user_id="u1", agent_id="default_agent",
|
|
client_meta=meta)
|
|
_run(writer._reconcile_and_store(
|
|
new_memory_texts=["用户用 React 18 + TS。"],
|
|
new_memories_meta=[{"content": "用户用 React 18 + TS。", "layer": "L2_FACT",
|
|
"tags": [], "owner": "user"}],
|
|
request=req, vector_store=store, req_id="rid-new",
|
|
))
|
|
|
|
def test_update_appends_to_existing_meta(self, tmp_path):
|
|
cfg, store = self._store(tmp_path)
|
|
node = self._seed(store, {"request_id": "rid-old", "meta": [META1]})
|
|
writer = self._writer(cfg, store, [self._update_op(node.node_id)])
|
|
|
|
self._write(writer, store, None, META2)
|
|
|
|
got = _run(store.get_by_id(node.node_id))
|
|
assert got.custom.get("meta") == [META1, META2], (
|
|
f"旧 meta 必须保留并追加新的，实际 {got.custom.get('meta')}"
|
|
)
|
|
assert got.content == "用户用 React 18 + TS。", "记忆本体仍需正常更新"
|
|
|
|
def test_update_without_meta_leaves_custom_untouched(self, tmp_path):
|
|
"""不传 meta 时不得触碰 custom——这是对存量写入路径的零影响保证。"""
|
|
cfg, store = self._store(tmp_path)
|
|
node = self._seed(store, {"request_id": "rid-old", "meta": [META1]})
|
|
writer = self._writer(cfg, store, [self._update_op(node.node_id)])
|
|
|
|
self._write(writer, store, None, None)
|
|
|
|
got = _run(store.get_by_id(node.node_id))
|
|
assert got.custom.get("meta") == [META1], "没传 meta 却改动了已有 meta"
|
|
assert got.custom.get("request_id") == "rid-old"
|
|
|
|
def test_update_on_node_without_meta(self, tmp_path):
|
|
cfg, store = self._store(tmp_path)
|
|
node = self._seed(store, {"request_id": "rid-old"})
|
|
writer = self._writer(cfg, store, [self._update_op(node.node_id)])
|
|
|
|
self._write(writer, store, None, META1)
|
|
|
|
got = _run(store.get_by_id(node.node_id))
|
|
assert got.custom.get("meta") == [META1]
|
|
assert got.custom.get("request_id") == "rid-old", "内部字段不能被合并冲掉"
|
|
|
|
def test_add_carries_single_meta(self, tmp_path):
|
|
from hy_memory.agent.reconciler import ReconcileOp
|
|
|
|
cfg, store = self._store(tmp_path)
|
|
op = ReconcileOp(op="ADD", content="用户用 Vue。", owner="user",
|
|
tags=[], layer="L2_FACT")
|
|
writer = self._writer(cfg, store, [op])
|
|
|
|
self._write(writer, store, None, META1)
|
|
|
|
nodes = _run(store.list_by_user(user_id="u1"))
|
|
added = [n for n in nodes if n.content == "用户用 Vue。"]
|
|
assert len(added) == 1
|
|
assert added[0].custom.get("meta") == [META1]
|
|
|
|
|
|
class TestReadPathsExposeMeta:
|
|
"""list / search / get 三处都要在顶层给 meta，读侧才不用为不同接口写两套。"""
|
|
|
|
def test_list_item_exposes_meta(self):
|
|
from hy_memory.models.memory import MemoryNode, MemoryLayer
|
|
|
|
node = MemoryNode(
|
|
node_id="n1", user_id="u1", layer=MemoryLayer.L2_FACT, content="c",
|
|
custom={"request_id": "rid", "meta": [META1]},
|
|
)
|
|
item = HyMemoryClient._memory_node_to_list_item(node)
|
|
assert item["meta"] == [META1]
|
|
|
|
def test_list_item_meta_empty_when_absent(self):
|
|
from hy_memory.models.memory import MemoryNode, MemoryLayer
|
|
|
|
node = MemoryNode(node_id="n1", user_id="u1", layer=MemoryLayer.L2_FACT, content="c")
|
|
assert HyMemoryClient._memory_node_to_list_item(node)["meta"] == []
|
|
|
|
def test_readers_expose_meta_from_custom(self):
|
|
"""两个 reader 的共享主流程必须透出 meta。"""
|
|
import hy_memory.pipelines.reader_base as reader_base
|
|
|
|
src = inspect.getsource(reader_base)
|
|
assert '"meta"' in src and 'get("meta")' in src
|
|
|
|
def test_readers_do_not_leak_whole_custom(self):
|
|
"""custom 含 request_id 等内部字段，不能整份塞进 search 返回项。"""
|
|
import hy_memory.pipelines.reader_base as reader_base
|
|
|
|
src = inspect.getsource(reader_base)
|
|
assert '"custom": ' not in src
|