Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 11 additions & 12 deletions src/memos/memories/textual/tree_text_memory/retrieve/recall.py
Original file line number Diff line number Diff line change
Expand Up @@ -367,7 +367,12 @@ def process_node(node):
final_nodes = [result for result in temp_results if result is not None]
return final_nodes

def _vector_recall(
def _vector_recall(self, *args, **kwargs) -> list[TextualMemoryItem]:
return [
TextualMemoryItem.from_dict(node) for node in self._vector_recall_nodes(*args, **kwargs)
]

def _vector_recall_nodes(
self,
query_embedding: list[list[float]],
memory_scope: str,
Expand All @@ -378,7 +383,7 @@ def _vector_recall(
search_filter: dict | None = None,
search_priority: dict | None = None,
user_name: str | None = None,
) -> list[TextualMemoryItem]:
) -> list[dict]:
"""
Perform vector-based similarity retrieval using query embedding.
# TODO: tackle with post-filter and pre-filter(5.18+) better.
Expand Down Expand Up @@ -477,7 +482,7 @@ def search_path_b():
hit.get("memory") is not None and hit.get("memory_type") is not None
for hit in lightweight_hits
):
return [self._memory_item_from_vector_hit(hit) for hit in lightweight_hits]
return [self._node_from_vector_hit(hit) for hit in lightweight_hits]
logger.warning(
"Lightweight vector search returned incomplete fields; falling back to get_nodes"
)
Expand Down Expand Up @@ -513,23 +518,17 @@ def search_path_b():
node["metadata"]["relativity"] = id_to_hit.get(rid, {}).get("score", 0.0)
ordered_nodes.append(node)

return [TextualMemoryItem.from_dict(n) for n in ordered_nodes]
return ordered_nodes

@staticmethod
def _memory_item_from_vector_hit(hit: dict) -> TextualMemoryItem:
def _node_from_vector_hit(hit: dict) -> dict:
metadata = {
key: value
for key, value in hit.items()
if key not in {"id", "memory", "score"} and value is not None
}
metadata["relativity"] = hit.get("score", 0.0)
return TextualMemoryItem.from_dict(
{
"id": hit["id"],
"memory": hit["memory"],
"metadata": metadata,
}
)
return {"id": hit["id"], "memory": hit["memory"], "metadata": metadata}

def _bm25_recall(
self,
Expand Down
12 changes: 7 additions & 5 deletions src/memos/memories/textual/tree_text_memory/retrieve/searcher.py
Original file line number Diff line number Diff line change
Expand Up @@ -1113,15 +1113,15 @@ def _retrieve_simple(
logger.info(f"[SIMPLESEARCH] Query words: {query_words}")
query_embeddings = self.embedder.embed(query_words)

items = self.graph_retriever.retrieve_from_mixed(
candidate_nodes = self.graph_retriever._vector_recall_nodes(
top_k=top_k * 2,
memory_scope=None,
query_embedding=query_embeddings,
search_filter=search_filter,
user_name=user_name,
)
logger.info(f"[SIMPLESEARCH] Items count: {len(items)}")
documents = [getattr(item, "memory", "") for item in items]
logger.info(f"[SIMPLESEARCH] Items count: {len(candidate_nodes)}")
documents = [node["memory"] for node in candidate_nodes]
if not documents:
return []
documents_embeddings = self.embedder.embed(documents)
Expand All @@ -1130,11 +1130,13 @@ def _retrieve_simple(
return []
similarity_matrix = cosine_similarity_matrix(documents_embeddings)
selected_indices, _ = find_best_unrelated_subgroup(documents, similarity_matrix)
selected_items = [items[i] for i in selected_indices]
rerank = bool(kwargs.get("rerank", True))
if not rerank or self.reranker is None:
selected_indices = selected_indices[:top_k]
selected_items = [TextualMemoryItem.from_dict(candidate_nodes[i]) for i in selected_indices]
logger.info(
f"[SIMPLESEARCH] after unrelated subgroup selection items count: {len(selected_items)}"
)
rerank = bool(kwargs.get("rerank", True))
return self._maybe_rerank(
rerank,
query=query,
Expand Down
11 changes: 9 additions & 2 deletions tests/graph_dbs/test_search_return_fields.py
Original file line number Diff line number Diff line change
Expand Up @@ -276,7 +276,7 @@ def test_polardb_declares_lightweight_vector_search_support(self, polardb_instan
],
)
def test_lightweight_numeric_user_id_can_be_recalled(
self, polardb_instance, stored_user_id, expected_user_id
self, polardb_instance, stored_user_id, expected_user_id, monkeypatch
):
from memos.memories.textual.tree_text_memory.retrieve.recall import GraphMemoryRetriever

Expand Down Expand Up @@ -324,7 +324,14 @@ def fake_connection():
"is_fast",
],
)
memory = GraphMemoryRetriever._memory_item_from_vector_hit(hits[0])
monkeypatch.setenv("MEMOS_POLARDB_LIGHTWEIGHT_SEARCH_ENABLED", "true")
graph_store = MagicMock(supports_lightweight_vector_search=True)
graph_store.search_by_embedding.return_value = hits
retriever = GraphMemoryRetriever(graph_store, MagicMock())
[memory] = retriever.retrieve_from_mixed(
top_k=1, query_embedding=[[0.1] * 1024], user_name="knowledge-cube"
)
graph_store.get_nodes.assert_not_called()
assert memory.metadata.user_id == expected_user_id
assert memory.metadata.tags == ["synthetic"]
assert memory.metadata.info == {"test": True}
Expand Down
69 changes: 69 additions & 0 deletions tests/memories/textual/test_deferred_candidate.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,69 @@
import uuid

from unittest.mock import Mock, patch

import pytest

from memos.memories.textual.item import TextualMemoryItem
from memos.memories.textual.tree_text_memory.retrieve.recall import GraphMemoryRetriever
from memos.memories.textual.tree_text_memory.retrieve.retrieval_mid_structs import ParsedTaskGoal
from memos.memories.textual.tree_text_memory.retrieve.searcher import Searcher


@pytest.fixture
def retriever():
store = Mock()
store.get_nodes.return_value = [
{
"id": str(uuid.uuid4()),
"memory": "first",
"metadata": {"metadata": {"memory_type": "SkillMemory"}},
},
{"id": str(uuid.uuid4()), "memory": "second", "metadata": {"memory_type": "UserMemory"}},
]
store.search_by_embedding.return_value = [
{"id": node["id"], "score": 1.0 - i / 10}
for i, node in enumerate(store.get_nodes.return_value)
]
return GraphMemoryRetriever(store, Mock())


@pytest.mark.parametrize("mixed", [False, True])
def test_public_recall_materializes_legacy_nested_metadata(retriever, mixed):
if mixed:
results = retriever.retrieve_from_mixed(top_k=2, query_embedding=[[0.1]])
else:
results = retriever.retrieve(
"query",
ParsedTaskGoal(keys=[], tags=[]),
top_k=2,
memory_scope="SkillMemory",
query_embedding=[[0.1]],
)

assert all(isinstance(item, TextualMemoryItem) for item in results)
assert [item.metadata.memory_type for item in results] == ["SkillMemory", "UserMemory"]


def test_plugin_search_materializes_only_subgroup_survivors(retriever):
searcher = Searcher.__new__(Searcher)
searcher.graph_retriever = retriever
searcher.tokenizer = None
searcher.embedder = Mock()
searcher.embedder.embed.return_value = [[0.1], [0.2]]
searcher.reranker = Mock()
searcher.reranker.rerank.side_effect = lambda **kw: [(kw["graph_results"][0], 0.7)]

with (
patch.object(TextualMemoryItem, "from_dict", wraps=TextualMemoryItem.from_dict) as build,
patch(
"memos.memories.textual.tree_text_memory.retrieve.searcher.find_best_unrelated_subgroup",
return_value=([0], None),
),
):
[(result, score)] = searcher._retrieve_simple("sample", top_k=1)

assert result.metadata.memory_type == "SkillMemory"
assert score == 0.7
build.assert_called_once()
assert build.call_args.args[0]["memory"] == "first"
Loading