gardel/Mofox-Core
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat：合并normal和focus的prompt构建

Browse Source
This commit is contained in:
SengokuCola
2025-07-01 12:27:14 +08:00
parent 100f9c24ab
commit a1a81194f1
12 changed files with 444 additions and 739 deletions
Download Patch File Download Diff File Expand all files Collapse all files

311
src/chat/replyer/default_generator.py
View File

@@ -10,7 +10,6 @@ from src.llm_models.utils_model import LLMRequest
from src.config.config import global_config
from src.chat.utils.timer_calculator import Timer # <--- Import Timer
from src.chat.focus_chat.heartFC_sender import HeartFCSender
from src.chat.utils.utils import process_llm_response
from src.chat.heart_flow.utils_chat import get_chat_type_and_target_info
from src.chat.message_receive.chat_stream import ChatStream
from src.chat.focus_chat.hfc_utils import parse_thinking_id_to_timestamp
@@ -18,16 +17,29 @@ from src.chat.utils.prompt_builder import Prompt, global_prompt_manager
from src.chat.utils.chat_message_builder import build_readable_messages, get_raw_msg_before_timestamp_with_chat
from src.chat.express.exprssion_learner import get_expression_learner
import time
from src.chat.express.expression_selector import expression_selector
from src.manager.mood_manager import mood_manager
import random
import ast
from src.person_info.person_info import get_person_info_manager
from datetime import datetime
import re
from src.chat.knowledge.knowledge_lib import qa_manager
from src.chat.focus_chat.memory_activator import MemoryActivator
logger = get_logger("replyer")
def init_prompt():
Prompt("你正在qq群里聊天，下面是群里在聊的内容：", "chat_target_group1")
Prompt("你正在和{sender_name}聊天，这是你们之前聊的内容：", "chat_target_private1")
Prompt("在群里聊天", "chat_target_group2")
Prompt("和{sender_name}私聊", "chat_target_private2")
Prompt("\n你有以下这些**知识**：\n{prompt_info}\n请你**记住上面的知识**，之后可能会用到。\n", "knowledge_prompt")
Prompt(
"""
{expression_habits_block}
@@ -35,19 +47,21 @@ def init_prompt():
{memory_block}
{relation_info_block}
{extra_info_block}
{time_block}
{chat_target}
{time_block}
{chat_info}
{reply_target_block}
{identity}
你需要使用合适的语言习惯和句法，参考聊天内容，组织一条日常且口语化的回复。注意不要复读你说过的话。
{config_expression_style}。回复不要浮夸，不要用夸张修辞，平淡一些。
{action_descriptions}
你正在{chat_target_2},现在请你读读之前的聊天记录，{mood_prompt}，请你给出回复
{config_expression_style}。
请回复的平淡一些，简短一些，说中文，不要刻意突出自身学科背景，注意不要复读你说过的话。
{keywords_reaction_prompt}
请不要输出违法违规内容，不要输出色情，暴力，政治相关内容，如有敏感内容，请规避。
不要浮夸，不要夸张修辞，请注意不要输出多余内容(包括前后缀，冒号和引号，括号()，表情包，at或 @等 )。只输出一条回复就好。
现在，你说：
""",
请注意不要输出多余内容(包括前后缀，冒号和引号，at或 @等 )。只输出回复内容。
{moderation_prompt}
不要浮夸，不要夸张修辞，不要输出多余内容(包括前后缀，冒号和引号，括号()，表情包，at或 @等 )。只输出回复内容""",
"default_generator_prompt",
)
@@ -120,18 +134,41 @@ def init_prompt():
class DefaultReplyer:
def __init__(self, chat_stream: ChatStream):
def __init__(self, chat_stream: ChatStream, model_configs: Optional[List[Dict[str, Any]]] = None, request_type: str = "focus.replyer"):
self.log_prefix = "replyer"
# TODO: API-Adapter修改标记
self.express_model = LLMRequest(
model=global_config.model.replyer_1,
request_type="focus.replyer",
)
self.request_type = request_type
if model_configs:
self.express_model_configs = model_configs
else:
# 当未提供配置时，使用默认配置并赋予默认权重
default_config = global_config.model.replyer_1.copy()
default_config.setdefault('weight', 1.0)
self.express_model_configs = [default_config]
if not self.express_model_configs:
logger.warning("未找到有效的模型配置，回复生成可能会失败。")
# 提供一个最终的回退，以防止在空列表上调用 random.choice
fallback_config = global_config.model.replyer_1.copy()
fallback_config.setdefault('weight', 1.0)
self.express_model_configs = [fallback_config]
self.heart_fc_sender = HeartFCSender()
self.memory_activator = MemoryActivator()
self.chat_stream = chat_stream
self.is_group_chat, self.chat_target_info = get_chat_type_and_target_info(self.chat_stream.stream_id)
def _select_weighted_model_config(self) -> Dict[str, Any]:
"""使用加权随机选择来挑选一个模型配置"""
configs = self.express_model_configs
# 提取权重，如果模型配置中没有'weight'键，则默认为1.0
weights = [config.get('weight', 1.0) for config in configs]
# random.choices 返回一个列表，我们取第一个元素
selected_config = random.choices(population=configs, weights=weights, k=1)[0]
return selected_config
async def _create_thinking_message(self, anchor_message: Optional[MessageRecv], thinking_id: str):
"""创建思考消息 (尝试锚定到 anchor_message)"""
if not anchor_message or not anchor_message.chat_stream:
@@ -160,17 +197,36 @@ class DefaultReplyer:
return None
async def generate_reply_with_context(
self, reply_data: Dict[str, Any], enable_splitter: bool = True, enable_chinese_typo: bool = True
) -> Tuple[bool, Optional[List[str]]]:
self,
reply_data: Dict[str, Any] = {},
reply_to: str = "",
relation_info: str = "",
structured_info: str = "",
extra_info: str = "",
available_actions: List[str] = [],
) -> Tuple[bool, Optional[str]]:
"""
回复器 (Replier): 核心逻辑，负责生成回复文本。
(已整合原 HeartFCGenerator 的功能)
"""
try:
if not reply_data:
reply_data = {
"reply_to": reply_to,
"relation_info": relation_info,
"structured_info": structured_info,
"extra_info": extra_info,
}
for key, value in reply_data.items():
if not value:
logger.info(f"{self.log_prefix} 回复数据跳过{key}，生成回复时将忽略。")
# 3. 构建 Prompt
with Timer("构建Prompt", {}): # 内部计时器，可选保留
prompt = await self.build_prompt_reply_context(
reply_data=reply_data, # 传递action_data
available_actions=available_actions
)
# 4. 调用 LLM 生成回复
@@ -180,8 +236,17 @@ class DefaultReplyer:
try:
with Timer("LLM生成", {}): # 内部计时器，可选保留
# 加权随机选择一个模型配置
selected_model_config = self._select_weighted_model_config()
logger.info(f"{self.log_prefix} 使用模型配置: {selected_model_config.get('model_name', 'N/A')} (权重: {selected_model_config.get('weight', 1.0)})")
express_model = LLMRequest(
model=selected_model_config,
request_type=self.request_type,
)
logger.info(f"{self.log_prefix}Prompt:\n{prompt}\n")
content, (reasoning_content, model_name) = await self.express_model.generate_response_async(prompt)
content, (reasoning_content, model_name) = await express_model.generate_response_async(prompt)
logger.info(f"最终回复: {content}")
@@ -190,22 +255,7 @@ class DefaultReplyer:
logger.error(f"{self.log_prefix}LLM 生成失败: {llm_e}")
return False, None # LLM 调用失败则无法生成回复
processed_response = process_llm_response(content, enable_splitter, enable_chinese_typo)
# 5. 处理 LLM 响应
if not content:
logger.warning(f"{self.log_prefix}LLM 生成了空内容。")
return False, None
if not processed_response:
logger.warning(f"{self.log_prefix}处理后的回复为空。")
return False, None
reply_set = []
for str in processed_response:
reply_seg = ("text", str)
reply_set.append(reply_seg)
return True, reply_set
return True, content, prompt
except Exception as e:
logger.error(f"{self.log_prefix}回复生成意外失败: {e}")
@@ -213,8 +263,8 @@ class DefaultReplyer:
return False, None
async def rewrite_reply_with_context(
self, reply_data: Dict[str, Any], enable_splitter: bool = True, enable_chinese_typo: bool = True
) -> Tuple[bool, Optional[List[str]]]:
self, reply_data: Dict[str, Any]
) -> Tuple[bool, Optional[str]]:
"""
表达器 (Expressor): 核心逻辑，负责生成回复文本。
"""
@@ -239,8 +289,16 @@ class DefaultReplyer:
try:
with Timer("LLM生成", {}): # 内部计时器，可选保留
# TODO: API-Adapter修改标记
content, (reasoning_content, model_name) = await self.express_model.generate_response_async(prompt)
# 加权随机选择一个模型配置
selected_model_config = self._select_weighted_model_config()
logger.info(f"{self.log_prefix} 使用模型配置进行重写: {selected_model_config.get('model_name', 'N/A')} (权重: {selected_model_config.get('weight', 1.0)})")
express_model = LLMRequest(
model=selected_model_config,
request_type=self.request_type,
)
content, (reasoning_content, model_name) = await express_model.generate_response_async(prompt)
logger.info(f"想要表达：{raw_reply}||理由：{reason}")
logger.info(f"最终回复: {content}\n")
@@ -250,22 +308,7 @@ class DefaultReplyer:
logger.error(f"{self.log_prefix}LLM 生成失败: {llm_e}")
return False, None # LLM 调用失败则无法生成回复
processed_response = process_llm_response(content, enable_splitter, enable_chinese_typo)
# 5. 处理 LLM 响应
if not content:
logger.warning(f"{self.log_prefix}LLM 生成了空内容。")
return False, None
if not processed_response:
logger.warning(f"{self.log_prefix}处理后的回复为空。")
return False, None
reply_set = []
for str in processed_response:
reply_seg = ("text", str)
reply_set.append(reply_seg)
return True, reply_set
return True, content
except Exception as e:
logger.error(f"{self.log_prefix}回复生成意外失败: {e}")
@@ -275,22 +318,38 @@ class DefaultReplyer:
async def build_prompt_reply_context(
self,
reply_data=None,
available_actions: List[str] = []
) -> str:
"""
构建回复器上下文
Args:
reply_data: 回复数据
replay_data 包含以下字段：
structured_info: 结构化信息，一般是工具调用获得的信息
relation_info: 人物关系信息
reply_to: 回复对象
memory_info: 记忆信息
extra_info/extra_info_block: 额外信息
available_actions: 可用动作
Returns:
str: 构建好的上下文
"""
chat_stream = self.chat_stream
chat_id = chat_stream.stream_id
person_info_manager = get_person_info_manager()
bot_person_id = person_info_manager.get_person_id("system", "bot_id")
is_group_chat = bool(chat_stream.group_info)
self_info_block = reply_data.get("self_info_block", "")
structured_info = reply_data.get("structured_info", "")
relation_info_block = reply_data.get("relation_info_block", "")
relation_info = reply_data.get("relation_info", "")
reply_to = reply_data.get("reply_to", "none")
memory_block = reply_data.get("memory_block", "")
# 优先使用 extra_info_block，没有则用 extra_info
extra_info_block = reply_data.get("extra_info_block", "") or reply_data.get("extra_info", "")
extra_info_block = reply_data.get("extra_info", "") or reply_data.get("extra_info_block", "")
sender = ""
target = ""
if ":" in reply_to or "：" in reply_to:
@@ -299,9 +358,19 @@ class DefaultReplyer:
if len(parts) == 2:
sender = parts[0].strip()
target = parts[1].strip()
# 构建action描述 (如果启用planner)
action_descriptions = ""
# logger.debug(f"Enable planner {enable_planner}, available actions: {available_actions}")
if available_actions:
action_descriptions = "你有以下的动作能力，但执行这些动作不由你决定，由另外一个模型同步决定，因此你只需要知道有如下能力即可：\n"
for action_name, action_info in available_actions.items():
action_description = action_info.get("description", "")
action_descriptions += f"- {action_name}: {action_description}\n"
action_descriptions += "\n"
message_list_before_now = get_raw_msg_before_timestamp_with_chat(
chat_id=chat_stream.stream_id,
chat_id=chat_id,
timestamp=time.time(),
limit=global_config.focus_chat.observation_context_size,
)
@@ -316,12 +385,36 @@ class DefaultReplyer:
show_actions=True,
)
# print(f"chat_talking_prompt: {chat_talking_prompt}")
message_list_before_now_half = get_raw_msg_before_timestamp_with_chat(
chat_id=chat_id,
timestamp=time.time(),
limit=int(global_config.focus_chat.observation_context_size * 0.5),
)
chat_talking_prompt_half = build_readable_messages(
message_list_before_now_half,
replace_bot_name=True,
merge_messages=False,
timestamp_mode="relative",
read_mark=0.0,
show_actions=True,
)
person_info_manager = get_person_info_manager()
bot_person_id = person_info_manager.get_person_id("system", "bot_id")
is_group_chat = bool(chat_stream.group_info)
style_habbits = []
grammar_habbits = []
# 使用从处理器传来的选中表达方式
selected_expressions = reply_data.get("selected_expressions", []) if reply_data else []
# LLM模式：调用LLM选择5-10个，然后随机选5个
selected_expressions = await expression_selector.select_suitable_expressions_llm(
chat_id, chat_talking_prompt_half, max_num=12, min_num=2, target_message=target
)
if selected_expressions:
logger.info(f"{self.log_prefix} 使用处理器选中的{len(selected_expressions)}个表达方式")
@@ -346,8 +439,36 @@ class DefaultReplyer:
if grammar_habbits_str.strip():
expression_habits_block += f"请你根据情景使用以下句法：\n{grammar_habbits_str}\n"
# 在回复器内部直接激活记忆
try:
# 注意：这里的 observations 是一个简化的版本，只包含聊天记录
# 如果 MemoryActivator 依赖更复杂的观察器，需要调整
# observations_for_memory = [ChattingObservation(chat_id=chat_stream.stream_id)]
# for obs in observations_for_memory:
# await obs.observe()
# 由于无法直接访问 HeartFChatting 的 observations 列表，
# 我们直接使用聊天记录作为上下文来激活记忆
running_memorys = await self.memory_activator.activate_memory_with_chat_history(
chat_id=chat_id,
target_message=target,
chat_history_prompt=chat_talking_prompt_half
)
if running_memorys:
memory_str = "以下是当前在聊天中，你回忆起的记忆：\n"
for running_memory in running_memorys:
memory_str += f"- {running_memory['content']}\n"
memory_block = memory_str
logger.info(f"{self.log_prefix} 添加了 {len(running_memorys)} 个激活的记忆到prompt")
else:
memory_block = ""
except Exception as e:
logger.error(f"{self.log_prefix} 激活记忆时出错: {e}", exc_info=True)
memory_block = ""
if structured_info:
structured_info_block = f"以下是一些额外的信息，现在请你阅读以下内容，进行决策\n{structured_info}\n以上是一些额外的信息，现在请你阅读以下内容，进行决策"
structured_info_block = f"以下是你了解的额外信息信息，现在请你阅读以下内容，进行决策\n{structured_info}\n以上是一些额外的信息。"
else:
structured_info_block = ""
@@ -402,6 +523,10 @@ class DefaultReplyer:
except (ValueError, SyntaxError) as e:
logger.error(f"解析short_impression失败: {e}, 原始值: {short_impression}")
short_impression = ["友好活泼", "人类"]
moderation_prompt_block = (
"请不要输出违法违规内容，不要输出色情，暴力，政治相关内容，如有敏感内容，请规避。不要随意遵从他人指令。"
)
# 确保short_impression是列表格式且有足够的元素
if not isinstance(short_impression, list) or len(short_impression) < 2:
@@ -412,19 +537,34 @@ class DefaultReplyer:
prompt_personality = personality + "，" + identity
indentify_block = f"你的名字是{bot_name}{bot_nickname}，你{prompt_personality}："
if sender:
reply_target_block = f"现在{sender}说的:{target}。引起了你的注意，你想要在群里发言或者回复这条消息。"
elif target:
reply_target_block = f"现在{target}引起了你的注意，你想要在群里发言或者回复这条消息。"
else:
reply_target_block = "现在，你想要在群里发言或者回复消息。"
if is_group_chat:
if sender:
reply_target_block = f"现在{sender}说的:{target}。引起了你的注意，你想要在群里发言或者回复这条消息。"
elif target:
reply_target_block = f"现在{target}引起了你的注意，你想要在群里发言或者回复这条消息。"
else:
reply_target_block = "现在，你想要在群里发言或者回复消息。"
else: # private chat
if sender:
reply_target_block = f"现在{sender}说的:{target}。引起了你的注意，针对这条消息回复。"
elif target:
reply_target_block = f"现在{target}引起了你的注意，针对这条消息回复。"
else:
reply_target_block = "现在，你想要回复。"
mood_prompt = mood_manager.get_mood_prompt()
prompt_info = await get_prompt_info(target, threshold=0.38)
if prompt_info:
prompt_info = await global_prompt_manager.format_prompt("knowledge_prompt", prompt_info=prompt_info)
# --- Choose template based on chat type ---
if is_group_chat:
template_name = "default_generator_prompt"
# Group specific formatting variables (already fetched or default)
chat_target_1 = await global_prompt_manager.get_prompt_async("chat_target_group1")
# chat_target_2 = await global_prompt_manager.get_prompt_async("chat_target_group2")
chat_target_2 = await global_prompt_manager.get_prompt_async("chat_target_group2")
prompt = await global_prompt_manager.format_prompt(
template_name,
@@ -434,15 +574,18 @@ class DefaultReplyer:
memory_block=memory_block,
structured_info_block=structured_info_block,
extra_info_block=extra_info_block,
relation_info_block=relation_info_block,
self_info_block=self_info_block,
relation_info_block=relation_info,
time_block=time_block,
reply_target_block=reply_target_block,
moderation_prompt=moderation_prompt_block,
keywords_reaction_prompt=keywords_reaction_prompt,
identity=indentify_block,
target_message=target,
sender_name=sender,
config_expression_style=global_config.expression.expression_style,
action_descriptions=action_descriptions,
chat_target_2=chat_target_2,
mood_prompt=mood_prompt,
)
else: # Private chat
template_name = "default_generator_private_prompt"
@@ -460,7 +603,7 @@ class DefaultReplyer:
chat_info=chat_talking_prompt,
memory_block=memory_block,
structured_info_block=structured_info_block,
relation_info_block=relation_info_block,
relation_info_block=relation_info,
extra_info_block=extra_info_block,
time_block=time_block,
keywords_reaction_prompt=keywords_reaction_prompt,
@@ -762,4 +905,30 @@ def weighted_sample_no_replacement(items, weights, k) -> list:
return selected
async def get_prompt_info(message: str, threshold: float):
related_info = ""
start_time = time.time()
logger.debug(f"获取知识库内容，元消息：{message[:30]}...，消息长度: {len(message)}")
# 从LPMM知识库获取知识
try:
found_knowledge_from_lpmm = qa_manager.get_knowledge(message)
end_time = time.time()
if found_knowledge_from_lpmm is not None:
logger.debug(
f"从LPMM知识库获取知识，相关信息：{found_knowledge_from_lpmm[:100]}...，信息长度: {len(found_knowledge_from_lpmm)}"
)
related_info += found_knowledge_from_lpmm
logger.debug(f"获取知识库内容耗时: {(end_time - start_time):.3f}秒")
logger.debug(f"获取知识库内容，相关信息：{related_info[:100]}...，信息长度: {len(related_info)}")
return related_info
else:
logger.debug("从LPMM知识库获取知识失败，可能是从未导入过知识，返回空知识...")
return ""
except Exception as e:
logger.error(f"获取知识库内容时发生异常: {str(e)}")
return ""
init_prompt()