diff --git a/src/memos/api/config.py b/src/memos/api/config.py index d9db93c1a..7f61d54ac 100644 --- a/src/memos/api/config.py +++ b/src/memos/api/config.py @@ -427,7 +427,7 @@ def get_reader_config() -> dict[str, Any]: "config": { "chunk_type": os.getenv("MEM_READER_CHAT_CHUNK_TYPE", "default"), "chunk_length": int(os.getenv("MEM_READER_CHAT_CHUNK_TOKEN_SIZE", 1600)), - "chunk_session": int(os.getenv("MEM_READER_CHAT_CHUNK_SESS_SIZE", 20)), + "chunk_session": int(os.getenv("MEM_READER_CHAT_CHUNK_SESS_SIZE", 10)), "chunk_overlap": int(os.getenv("MEM_READER_CHAT_CHUNK_OVERLAP", 2)), }, } diff --git a/src/memos/configs/memory.py b/src/memos/configs/memory.py index 49320fbf5..34967849a 100644 --- a/src/memos/configs/memory.py +++ b/src/memos/configs/memory.py @@ -184,7 +184,7 @@ class TreeTextMemoryConfig(BaseTextMemoryConfig): ), ) - search_strategy: dict[str, bool] | None = Field( + search_strategy: dict[str, Any] | None = Field( default=None, description=( 'Set search strategy for this memory configuration.{"bm25": true, "cot": false}' diff --git a/src/memos/mem_reader/strategy_struct.py b/src/memos/mem_reader/strategy_struct.py index 2cac1652a..1fc21461e 100644 --- a/src/memos/mem_reader/strategy_struct.py +++ b/src/memos/mem_reader/strategy_struct.py @@ -43,7 +43,7 @@ def _get_llm_response(self, mem_str: str) -> dict: template = STRATEGY_PROMPT_DICT["chat"][lang] examples = STRATEGY_PROMPT_DICT["chat"][f"{lang}_example"] prompt = template.replace("${conversation}", mem_str) - if self.config.remove_prompt_example: + if self.config.remove_prompt_example: # TODO unused prompt = prompt.replace(examples, "") messages = [{"role": "user", "content": prompt}] try: @@ -112,6 +112,19 @@ def get_scene_data_info(self, scene_data: list, type: str) -> list[str]: results.append([overlap_item, item]) current_length = overlap_length + content_length + else: + cut_size, cut_overlap = ( + self.chat_chunker["chunk_session"], + self.chat_chunker["chunk_overlap"], + ) + for items in scene_data: + step = cut_size - cut_overlap + end = len(items) - cut_overlap + if end <= 0: + results.extend([items[:]]) + else: + results.extend([items[i : i + cut_size] for i in range(0, end, step)]) + elif type == "doc": parser_config = ParserConfigFactory.model_validate( { diff --git a/src/memos/memories/textual/simple_tree.py b/src/memos/memories/textual/simple_tree.py index 992b7bfab..313989cd2 100644 --- a/src/memos/memories/textual/simple_tree.py +++ b/src/memos/memories/textual/simple_tree.py @@ -66,7 +66,9 @@ def __init__( time_start_bm = time.time() self.search_strategy = config.search_strategy self.bm25_retriever = ( - EnhancedBM25() if self.search_strategy and self.search_strategy["bm25"] else None + EnhancedBM25() + if self.search_strategy and self.search_strategy.get("bm25", False) + else None ) logger.info(f"time init: bm25_retriever time is: {time.time() - time_start_bm}") diff --git a/src/memos/memories/textual/tree_text_memory/retrieve/retrieval_mid_structs.py b/src/memos/memories/textual/tree_text_memory/retrieve/retrieval_mid_structs.py index 6accc4a16..7aefaa1a3 100644 --- a/src/memos/memories/textual/tree_text_memory/retrieve/retrieval_mid_structs.py +++ b/src/memos/memories/textual/tree_text_memory/retrieve/retrieval_mid_structs.py @@ -13,3 +13,4 @@ class ParsedTaskGoal: rephrased_query: str | None = None internet_search: bool = False goal_type: str | None = None # e.g., 'default', 'explanation', etc. + context: str = "" diff --git a/src/memos/memories/textual/tree_text_memory/retrieve/retrieve_utils.py b/src/memos/memories/textual/tree_text_memory/retrieve/retrieve_utils.py index eec827c86..3f2b41a47 100644 --- a/src/memos/memories/textual/tree_text_memory/retrieve/retrieve_utils.py +++ b/src/memos/memories/textual/tree_text_memory/retrieve/retrieve_utils.py @@ -17,7 +17,7 @@ def find_project_root(marker=".git"): if (current / marker).exists(): return current current = current.parent - logger.warn(f"The project root directory tag file was not found: {marker}") + return Path(".") PROJECT_ROOT = find_project_root() diff --git a/src/memos/memories/textual/tree_text_memory/retrieve/searcher.py b/src/memos/memories/textual/tree_text_memory/retrieve/searcher.py index 0974d67f2..2f6ef6afa 100644 --- a/src/memos/memories/textual/tree_text_memory/retrieve/searcher.py +++ b/src/memos/memories/textual/tree_text_memory/retrieve/searcher.py @@ -30,8 +30,8 @@ logger = get_logger(__name__) COT_DICT = { - "fast": {"en": COT_PROMPT, "zh": COT_PROMPT_ZH}, - "fine": {"en": SIMPLE_COT_PROMPT, "zh": SIMPLE_COT_PROMPT_ZH}, + "fine": {"en": COT_PROMPT, "zh": COT_PROMPT_ZH}, + "fast": {"en": SIMPLE_COT_PROMPT, "zh": SIMPLE_COT_PROMPT_ZH}, } @@ -59,12 +59,8 @@ def __init__( # Create internet retriever from config if provided self.internet_retriever = internet_retriever self.moscube = moscube - self.vec_cot = ( - search_strategy.get("vec_cot", "false") == "true" if search_strategy else False - ) - self.use_fast_graph = ( - search_strategy.get("fast_graph", "false") == "true" if search_strategy else False - ) + self.vec_cot = search_strategy.get("cot", False) if search_strategy else False + self.use_fast_graph = search_strategy.get("fast_graph", False) if search_strategy else False self._usage_executor = ContextThreadPoolExecutor(max_workers=4, thread_name_prefix="usage") @@ -287,6 +283,7 @@ def _retrieve_paths( search_filter, user_name, id_filter, + mode=mode, ) ) tasks.append( @@ -369,6 +366,7 @@ def _retrieve_from_long_term_and_user( search_filter: dict | None = None, user_name: str | None = None, id_filter: dict | None = None, + mode: str = "fast", ): """Retrieve and rerank from LongTermMemory and UserMemory""" results = [] @@ -377,7 +375,7 @@ def _retrieve_from_long_term_and_user( # chain of thinking cot_embeddings = [] if self.vec_cot: - queries = self._cot_query(query) + queries = self._cot_query(query, mode=mode, context=parsed_goal.context) if len(queries) > 1: cot_embeddings = self.embedder.embed(queries) cot_embeddings.extend(query_embedding) @@ -566,7 +564,6 @@ def _cot_query( prompt = template.replace("${original_query}", query).replace( "${split_num_threshold}", str(split_num) ) - logger.info("COT process") messages = [{"role": "user", "content": prompt}] try: diff --git a/src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py b/src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py index 5d706559c..55e33494c 100644 --- a/src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py +++ b/src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py @@ -39,7 +39,7 @@ def parse( - mode == 'fine': use LLM to parse structured topic/keys/tags """ if mode == "fast": - return self._parse_fast(task_description, **kwargs) + return self._parse_fast(task_description, context=context, **kwargs) elif mode == "fine": if not self.llm: raise ValueError("LLM not provided for slow mode.") @@ -51,6 +51,7 @@ def _parse_fast(self, task_description: str, **kwargs) -> ParsedTaskGoal: """ Fast mode: simple jieba word split. """ + context = kwargs.get("context", "") use_fast_graph = kwargs.get("use_fast_graph", False) if use_fast_graph: desc_tokenized = self.tokenizer.tokenize_mixed(task_description) @@ -61,6 +62,7 @@ def _parse_fast(self, task_description: str, **kwargs) -> ParsedTaskGoal: goal_type="default", rephrased_query=task_description, internet_search=False, + context=context, ) else: return ParsedTaskGoal( @@ -70,6 +72,7 @@ def _parse_fast(self, task_description: str, **kwargs) -> ParsedTaskGoal: goal_type="default", rephrased_query=task_description, internet_search=False, + context=context, ) def _parse_fine( @@ -91,16 +94,17 @@ def _parse_fine( logger.info(f"Parsing Goal... LLM input is {prompt}") response = self.llm.generate(messages=[{"role": "user", "content": prompt}]) logger.info(f"Parsing Goal... LLM Response is {response}") - return self._parse_response(response) + return self._parse_response(response, context=context) except Exception: logger.warning(f"Fail to fine-parse query {query}: {traceback.format_exc()}") - return self._parse_fast(query) + return self._parse_fast(query, context=context) - def _parse_response(self, response: str) -> ParsedTaskGoal: + def _parse_response(self, response: str, **kwargs) -> ParsedTaskGoal: """ Parse LLM JSON output safely. """ try: + context = kwargs.get("context", "") response = response.replace("```", "").replace("json", "").strip() response_json = eval(response) return ParsedTaskGoal( @@ -110,6 +114,7 @@ def _parse_response(self, response: str) -> ParsedTaskGoal: rephrased_query=response_json.get("rephrased_instruction", None), internet_search=response_json.get("internet_search", False), goal_type=response_json.get("goal_type", "default"), + context=context, ) except Exception as e: raise ValueError(f"Failed to parse LLM output: {e}\nRaw response:\n{response}") from e diff --git a/src/memos/templates/mem_reader_strategy_prompts.py b/src/memos/templates/mem_reader_strategy_prompts.py index fca4d717b..ba4a00d0a 100644 --- a/src/memos/templates/mem_reader_strategy_prompts.py +++ b/src/memos/templates/mem_reader_strategy_prompts.py @@ -16,8 +16,13 @@ - Always set "model_type" to "UserMemory" for this output. 3. Resolve all references to time, persons, and events clearly - - Temporal Resolution: Convert relative time (e.g., 'yesterday') to absolute dates based on the message timestamp. Distinguish between event time and message time; flag any uncertainty. + - Temporal Resolution: Convert relative time (e.g., "yesterday") to absolute dates based on the message timestamp. Distinguish between event time and message time; flag any uncertainty. + > Where feasible, use the message timestamp to convert relative time expressions into absolute dates (e.g., "yesterday" in a message dated January 15, 2023, can be converted to "January 14, 2023," and "last week" can be described as "the week preceding January 15, 2023"). + > Explicitly differentiate between the time when the event occurred and the time the message was sent. + > Clearly indicate any uncertainty (e.g., "approximately June 2025", "exact date unknown"). - Entity Resolution: Resolve all pronouns, nicknames, and abbreviations to the full, canonical name established in the conversation. + > For example, "Melanie" uses the abbreviated name "Mel" in the paragraph; when extracting her name in the "value" field, it should be restored to "Melanie". + - Location resolution: If specific locations are mentioned, include them explicitly. 4. Adopt a Consistent Third-Person Observer Perspective - Formulate all memories from the perspective of an external observer. Use "The user" or their specific name as the subject. @@ -38,14 +43,14 @@ 7. Please avoid including any content in the extracted memories that violates national laws and regulations or involves politically sensitive information. -Return a valid JSON object with the following structure: +Return a valid JSON object with the following structure: { "memory list": [ { "key": , "memory_type": , - "value": , + "value": , "tags": }, ... @@ -54,11 +59,11 @@ } Language rules: -- The `key`, `value`, `tags`, and `summary` fields must match the primary language of the input conversation. **If the input is Chinese, output in Chinese.** -- Keep `memory_type` in English. +- The `key`, `value`, `tags`, `summary` and `memory_type` fields must be in English. + Example: -Conversation: +Conversations: user: [June 26, 2025 at 3:00 PM]: Hi Jerry! Yesterday at 3 PM I had a meeting with my team about the new project. assistant: Oh Tom! Do you think the team can finish by December 15? user: [June 26, 2025 at 3:00 PM]: I’m worried. The backend won’t be done until December 10, so testing will be tight. @@ -71,65 +76,19 @@ { "key": "Initial project meeting", "memory_type": "LongTermMemory", - "value": "[user-Tom viewpoint] On June 25, 2025 at 3:00 PM, Tom met with the team to discuss a new project. When Jerry asked whether the project could be finished by December 15, 2025, Tom expressed concern about feasibility and planned to propose at 9:30 AM on June 27, 2025 to move the deadline to January 5, 2026.", + "value": "[user-Tom viewpoint] On June 25, 2025 at 3:00 PM, Tom held a meeting with their team to discuss a new project. The conversation covered the timeline and raised concerns about the feasibility of the December 15, 2025 deadline.", "tags": ["Tom", "project", "timeline", "meeting", "deadline"] }, { - "key": "Jerry’s suggestion about the deadline", - "memory_type": "LongTermMemory", - "value": "[assistant-Jerry viewpoint] Jerry questioned the December 15 deadline and suggested considering an extension.", - "tags": ["Jerry", "deadline change", "suggestion"] + "key": "Planned scope adjustment", + "memory_type": "UserMemory", + "value": "Tom planned to suggest in a meeting on June 27, 2025 at 9:30 AM that the team should prioritize features and propose shifting the project deadline to January 5, 2026.", + "tags": ["Tom", "planning", "deadline change", "feature prioritization"] } ], - "summary": "Tom is currently working on a tight-schedule project. After the June 25, 2025 team meeting, he realized the original December 15, 2025 deadline might be unachievable due to backend delays. Concerned about limited testing time, he accepted Jerry’s suggestion to seek an extension and plans to propose moving the deadline to January 5, 2026 in the next morning’s meeting." + "summary": "Tom is currently focused on managing a new project with a tight schedule. After a team meeting on June 25, 2025, he realized the original deadline of December 15 might not be feasible due to backend delays. Concerned about insufficient testing time, he welcomed Jerry’s suggestion of proposing an extension. Tom plans to raise the idea of shifting the deadline to January 5, 2026 in the next morning’s meeting. His actions reflect both stress about timelines and a proactive, team-oriented problem-solving approach." } -Another Example in Chinese (Note: when the user's language is Chinese, you must also output in Chinese): - -对话（节选）： -user｜2025-07-12 19:30：下个月同事婚礼，要买条裙子，预算1500内，别太暴露。 -assistant｜19:32 -：了解～婚礼场合建议中长款，端庄一些。可以看法式茶歇/直筒/裹身三路。颜色上，奶油白、浅粉、雾霾蓝都适合；面料可看醋酸/真丝混纺（垂感好但需干洗）。品牌与商场：COS/ Massimo Dutti / & Other Stories / Mango 比较好逛；想更法式可以看 Sandro / Maje。 我记得你很喜欢Maje？ -user｜19:35：不喜欢亮色。国贸方便。 -assistant｜19:37：那亮度就压低些：雾霾蓝、藕粉、燕麦米白更稳。国贸有 COS / Massimo Dutti / &OS；可以先看 COS 直筒或裹身，再去 MD 看藕粉/雾霾蓝。搭配上，薄披肩/小西装都能压场。 -user｜19:40：165cm，S码；最好有口袋。 -assistant｜19:42：OK。版型避开大摆+高开衩，选直筒或轻A字更稳；若要口袋，COS直筒一些款有；开衩别太高，及小腿肚附近即可。鞋包：中跟裸色鞋 + 小号手拿包。 -user｜19:47：本来想茶歇A字，但雾霾蓝直筒也行。 -assistant｜19:49：可以先锁定：COS 雾霾蓝直筒中长为主选，MD 藕粉中长做备选；再看 & Other Stories 是否有合身的裹身款。 -user｜19:52：行，周六（7/19）去国贸试，合适就买。 -assistant｜19:54：记得看肩带可调/有无内衬，醋酸/真丝优先干洗；准备同色安全裤/防走光贴。如果当天没货，可下单调货或线上下单门店自提。 - -{ - "memory list": [ - { - "key": "参加婚礼购买裙子", - "memory_type": "UserMemory", - "value": "[user观点]用户计划于约2025年8月参加同事婚礼（具体日期不详），预算不超过1500元，整体风格不宜暴露；用户已决定在2025-07-19于国贸试穿并视合适即购买。", - "tags": ["婚礼", "预算", "国贸", "计划"] - }, - { - "key": "审美与版型偏好", - "memory_type": "UserMemory", - "value": "[user观点]用户不喜欢亮色，倾向低亮度色系；裙装偏好端庄的中长款，接受直筒或轻A字。", - "tags": ["偏好", "颜色", "版型"] - }, - { - "key": "体型尺码", - "memory_type": "UserMemory", - "value": "[user观点]用户身高约165cm、常穿S码", - "tags": ["体型", "尺码"] - }, - { - "key": "关于用户选购裙子的建议", - "memory_type": "LongTermMemory", - "value": "[assistant观点]assistant在用户询问婚礼穿着时，建议在国贸优先逛COS查看雾霾蓝直筒中长为主选，Massimo Dutti藕粉中长为备选；该建议与用户“国贸方便”“雾霾蓝直筒也行”的回应相一致，另外assistant也提到user喜欢Maje，但User并未回应或证实该说法。", - "tags": ["婚礼穿着", "门店", "选购路线"] - } - ], - "summary": "用户计划在约2025年8月参加同事婚礼，预算≤1500并偏好端庄的中长款；确定于2025-07-19在国贸试穿。其长期画像显示：不喜欢亮色、偏好低亮度色系与不过分暴露的版型，身高约165cm、S码且偏好裙装带口袋。助手提出的国贸选购路线以COS雾霾蓝直筒中长为主选、MD藕粉中长为备选，且与用户回应一致，为线下试穿与购买提供了明确路径。" -} - -Always respond in the same language as the conversation. Conversation: ${conversation} @@ -155,7 +114,11 @@ 3. 明确解析所有指代关系 - 时间解析：根据消息时间戳将相对时间（如“昨天”）转换为绝对日期。区分事件时间与消息时间，对不确定项进行标注 + # 条件允许则使用消息时间戳将相对时间表达转换为绝对日期（如：2023年1月15日的“昨天”则转换为2023年1月14日）；“上周”则转换为2023年1月15日前一周）。 + # 明确区分事件时间和消息时间。 + # 如果存在不确定性，需明确说明（例如，“约2025年6月”，“具体日期不详”）。 - 实体解析：将所有代词、昵称和缩写解析为对话中确立的完整规范名称 + - 地点解析：若提及具体地点，请包含在内。 4. 采用统一的第三人称观察视角 - 所有记忆表述均需从外部观察者视角构建，使用“用户”或其具体姓名作为主语 @@ -183,7 +146,7 @@ { "key": <字符串，唯一且简洁的记忆标题>, "memory_type": <字符串，"LongTermMemory" 或 "UserMemory">, - "value": <详细、独立且无歧义的记忆陈述——若输入对话为英文，则用英文；若为中文，则用中文>, + "value": <详细、独立且无歧义的记忆陈述>, "tags": <一个包含相关人名、事件和特征关键词的列表（例如，["丽丽","截止日期", "团队", "计划"]）> }, ... @@ -192,10 +155,10 @@ } 语言规则： -- `key`、`value`、`tags`、`summary` 字段必须与输入对话的主要语言一致。**如果输入是中文，请输出中文** -- `memory_type` 保持英文。 +- `key`、`value`、`tags`、`summary` 、`memory_type` 字段必须输出中文 -示例： + +示例1： 对话： user: [2025年6月26日下午3:00]：嗨Jerry！昨天下午3点我和团队开了个会，讨论新项目。 assistant: 哦Tom！你觉得团队能在12月15日前完成吗？ @@ -209,25 +172,20 @@ { "key": "项目初期会议", "memory_type": "LongTermMemory", - "value": "[user-Tom观点]2025年6月25日下午3:00，Tom与团队开会讨论新项目。当Jerry - 询问该项目能否在2025年12月15日前完成时，Tom对此日期前完成的可行性表达担忧，并计划在2025年6月27日上午9:30 - 提议将截止日期推迟至2026年1月5日。", - "tags": ["Tom", "项目", "时间表", "会议", "截止日期"] + "value": "2025年6月25日下午3:00，Tom与团队开会讨论新项目。会议涉及时间表，并提出了对2025年12月15日截止日期可行性的担忧。", + "tags": ["项目", "时间表", "会议", "截止日期"] }, { - "key": "Jerry对新项目截止日期的建议", - "memory_type": "LongTermMemory", - "value": "[assistant-Jerry观点]Jerry对Tom的新项目截止日期提出疑问、并提议Tom考虑延期。", - "tags": ["Jerry", "截止日期变更", "建议"] + "key": "计划调整范围", + "memory_type": "UserMemory", + "value": "Tom计划在2025年6月27日上午9:30的会议上建议团队优先处理功能，并提议将项目截止日期推迟至2026年1月5日。", + "tags": ["计划", "截止日期变更", "功能优先级"] } ], - "summary": "Tom目前正在做一个进度紧张的新项目。在2025年6月25日的团队会议后，他意识到原定2025年12月15 - 日的截止日期可能无法实现，因为后端会延迟。由于担心测试时间不足，他接受了Jerry提出的延期建议，计划在次日早上的会议上提出将截止日期推迟至2026 - 年1月5日。" + "summary": "Tom目前正专注于管理一个进度紧张的新项目。在2025年6月25日的团队会议后，他意识到原定2025年12月15日的截止日期可能无法实现，因为后端会延迟。由于担心测试时间不足，他接受了Jerry提出的延期建议。Tom计划在次日早上的会议上提出将截止日期推迟至2026年1月5日。他的行为反映出对时间线的担忧，以及积极、以团队为导向的问题解决方式。" } -另一个中文示例（注意：当用户语言为中文时，您也需输出中文）： - +示例2： 对话（节选）： user｜2025-07-12 19:30：下个月同事婚礼，要买条裙子，预算1500内，别太暴露。 assistant｜19:32 @@ -271,7 +229,6 @@ "summary": "用户计划在约2025年8月参加同事婚礼，预算≤1500并偏好端庄的中长款；确定于2025-07-19在国贸试穿。其长期画像显示：不喜欢亮色、偏好低亮度色系与不过分暴露的版型，身高约165cm、S码且偏好裙装带口袋。助手提出的国贸选购路线以COS雾霾蓝直筒中长为主选、MD藕粉中长为备选，且与用户回应一致，为线下试穿与购买提供了明确路径。" } -请始终使用与对话相同的语言进行回复。 对话： ${conversation}