From 0b90060288aa1b5f515716b98f3281e5bdeffc4a Mon Sep 17 00:00:00 2001 From: bittergreen Date: Tue, 28 Jul 2026 15:56:11 +0800 Subject: [PATCH] fix: preserve memory alongside source text in search results --- src/memos/api/handlers/formatters_handler.py | 11 ++++- tests/api/test_search_pipeline_hooks.py | 42 ++++++++++++++++++++ 2 files changed, 51 insertions(+), 2 deletions(-) diff --git a/src/memos/api/handlers/formatters_handler.py b/src/memos/api/handlers/formatters_handler.py index 638565bb9..ab205df2e 100644 --- a/src/memos/api/handlers/formatters_handler.py +++ b/src/memos/api/handlers/formatters_handler.py @@ -8,6 +8,7 @@ from typing import Any from memos.log import get_logger +from memos.memories.textual.tree_text_memory.retrieve.retrieve_utils import detect_lang from memos.templates.instruction_completion import instruct_completion @@ -196,9 +197,15 @@ def rerank_knowledge_mem( key=lambda item: item.get("metadata", {}).get("relativity", 0.0), reverse=True, ) - # replace memory value with source.content for LongTermMemory, WorkingMemory or UserMemory + # Combine the extracted memory with its original file source. + is_chinese_query = detect_lang(query) == "zh" for item in reranked_knowledge_mem: - item["memory"] = item["metadata"]["sources"][0]["content"] + memory = item.get("memory", "") + source_content = item["metadata"]["sources"][0]["content"] + if is_chinese_query: + item["memory"] = f"记忆:{memory},原文:{source_content}" + else: + item["memory"] = f"Memory: {memory}, Original text: {source_content}" item["metadata"]["sources"] = [] if strip_conversation_sources: diff --git a/tests/api/test_search_pipeline_hooks.py b/tests/api/test_search_pipeline_hooks.py index e96c3fbf5..befff6b80 100644 --- a/tests/api/test_search_pipeline_hooks.py +++ b/tests/api/test_search_pipeline_hooks.py @@ -15,6 +15,18 @@ def _memory(memory_id: str, memory: str, memory_type: str = "LongTermMemory") -> } +def _file_memory(memory_id: str, memory: str, source_content: str) -> dict: + return { + "id": memory_id, + "memory": memory, + "metadata": { + "memory_type": "LongTermMemory", + "relativity": 1.0, + "sources": [{"type": "file", "content": source_content}], + }, + } + + def test_search_request_passes_context_format_through_to_plugins(): req = APISearchRequest( user_id="user", @@ -76,3 +88,33 @@ def test_rerank_knowledge_mem_can_strip_conversation_sources(): conversation = next(item for item in reranked if item["memory"] == "conversation memory") assert conversation["metadata"]["sources"] == [] + + +def test_rerank_knowledge_mem_combines_memory_and_source_for_chinese_query(): + text_mem = [ + { + "cube_id": "cube", + "memories": [_file_memory("mem-1", "抽取后的记忆", "文件中的原文")], + } + ] + + reranked = rerank_knowledge_mem(None, "用户的中文问题", text_mem, top_k=1)[0]["memories"] + + assert reranked[0]["memory"] == "记忆:抽取后的记忆,原文:文件中的原文" + assert reranked[0]["metadata"]["sources"] == [] + + +def test_rerank_knowledge_mem_combines_memory_and_source_for_english_query(): + text_mem = [ + { + "cube_id": "cube", + "memories": [_file_memory("mem-1", "extracted memory", "original file text")], + } + ] + + reranked = rerank_knowledge_mem(None, "What does the file say?", text_mem, top_k=1)[0][ + "memories" + ] + + assert reranked[0]["memory"] == ("Memory: extracted memory, Original text: original file text") + assert reranked[0]["metadata"]["sources"] == []