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
36 changes: 35 additions & 1 deletion src/memos/mem_reader/multi_modal_struct.py
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
import traceback
import uuid

from collections.abc import Hashable
from typing import TYPE_CHECKING, Any

from memos import log
Expand All @@ -15,7 +16,11 @@
from memos.mem_reader.read_skill_memory.process_skill_memory import process_skill_memory_fine
from memos.mem_reader.simple_struct import PROMPT_DICT, SimpleStructMemReader
from memos.mem_reader.utils import parse_json_result
from memos.memories.textual.item import TextualMemoryItem, TreeNodeTextualMemoryMetadata
from memos.memories.textual.item import (
SourceMessage,
TextualMemoryItem,
TreeNodeTextualMemoryMetadata,
)
from memos.plugins.hook_defs import H
from memos.plugins.hooks import trigger_hook, trigger_single_hook
from memos.templates.mem_reader_prompts import MEMORY_MERGE_PROMPT_EN, MEMORY_MERGE_PROMPT_ZH
Expand All @@ -31,6 +36,28 @@
logger = log.get_logger(__name__)


def _freeze_source_value(value: Any) -> Hashable:
"""Ignore mapping order and preserve sequence order in provenance."""
if isinstance(value, dict):
return dict, frozenset((key, _freeze_source_value(item)) for key, item in value.items())
if isinstance(value, list | tuple):
return type(value), tuple(_freeze_source_value(item) for item in value)
return type(value), value


def _source_signature(source: SourceMessage | dict[str, Any]) -> Hashable:
"""Include all supplied provenance fields when deduplicating a window."""
try:
data = source.model_dump() if isinstance(source, SourceMessage) else source
signature = frozenset(
(key, _freeze_source_value(value)) for key, value in data.items() if value is not None
)
except (AttributeError, TypeError, ValueError, RecursionError):
# Keep opaque provenance distinct unless it is the same object.
return id(source)
return signature


class MultiModalStructMemReader(SimpleStructMemReader):
"""Multimodal implementation of MemReader that inherits from
SimpleStructMemReader."""
Expand Down Expand Up @@ -166,6 +193,7 @@ def _create_chunk_item(chunk_idx: int, chunk):
memory_type=item.metadata.memory_type,
tags=item.metadata.tags or [],
key=item.metadata.key,
# A chunk can land in a different window from the first chunk.
sources=item.metadata.sources or [],
background=item.metadata.background or "",
need_embed=False,
Expand Down Expand Up @@ -332,6 +360,8 @@ def _build_window_from_items(
# Collect all memory texts and sources
memory_texts = []
all_sources = []
# Keep one copy of each source per window while preserving its first position.
seen_source_sigs: set[Hashable] = set()
roles = set()
aggregated_file_ids: list[str] = []
ingest_batch_ids: set[str] = set()
Expand All @@ -346,6 +376,10 @@ def _build_window_from_items(
item_sources = [item_sources]

for source in item_sources:
sig = _source_signature(source)
if sig in seen_source_sigs:
continue
seen_source_sigs.add(sig)
# Add source to all_sources
all_sources.append(source)

Expand Down
89 changes: 89 additions & 0 deletions tests/mem_reader/test_sources_dedup.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,89 @@
from unittest.mock import Mock

import pytest

from memos.chunkers.simple_chunker import SimpleTextSplitter
from memos.mem_reader.multi_modal_struct import MultiModalStructMemReader
from memos.memories.textual.item import (
SourceMessage,
TextualMemoryItem,
TreeNodeTextualMemoryMetadata,
)


@pytest.fixture
def reader():
reader = MultiModalStructMemReader.__new__(MultiModalStructMemReader)
reader.chunker = SimpleTextSplitter(chunk_size=4, chunk_overlap=0)
reader._count_tokens = len
reader.embedder = Mock()
reader.embedder.embed.side_effect = lambda texts: [[0.1] for _ in texts]
return reader


def make_item(text: str, *sources: SourceMessage) -> TextualMemoryItem:
return TextualMemoryItem(
memory=text,
metadata=TreeNodeTextualMemoryMetadata(
user_id="test-user",
session_id="test-session",
memory_type="UserMemory",
sources=list(sources),
),
)


def test_window_deduplicates_nested_provenance_in_source_order(reader):
source = SourceMessage(
type="doc",
doc_path="sample.md",
content="A paragraph.",
file_info={"locator": {"pages": [1, 2], "offset": 4}},
span=[{"start": 4, "end": 16}],
)
duplicate = SourceMessage(
type="doc",
doc_path="sample.md",
content="A paragraph.",
file_info={"locator": {"offset": 4, "pages": [1, 2]}},
span=[{"end": 16, "start": 4}],
)
different_span = duplicate.model_copy(update={"span": [{"start": 20, "end": 32}]})
different_page_order = duplicate.model_copy(
update={"file_info": {"locator": {"pages": [2, 1], "offset": 4}}}
)

window = reader._build_window_from_items(
[
make_item("first", source),
make_item("second", duplicate, different_span, different_page_order),
],
{},
)

assert window.metadata.sources == [source, different_span, different_page_order]


def test_long_message_keeps_one_source_per_window(reader):
text = "aaaabbbbccccdddd"
source = SourceMessage(type="chat", role="user", content=text, message_id="msg-1")

windows = reader._concat_multi_modal_memories(
[make_item(text, source)], max_tokens=10, overlap=0
)

assert [window.memory for window in windows] == ["aaaabbbb", "ccccdddd"]
assert [window.metadata.sources for window in windows] == [[source], [source]]
assert [window.metadata.memory_type for window in windows] == ["UserMemory", "UserMemory"]


def test_window_keeps_identical_text_from_distinct_messages(reader):
first = SourceMessage(role="user", content="yes", message_id="msg-1")
second = SourceMessage(role="assistant", content="yes", message_id="msg-2")

window = reader._build_window_from_items(
[make_item("first", first), make_item("second", second, first.model_copy())], {}
)

assert window.metadata.sources == [first, second]
assert window.metadata.memory_type == "LongTermMemory"
Loading