better：优化做梦表现

2025-12-04 20:00:20 +08:00
parent 0399f878f0
commit add29efefc
12 changed files with 893 additions and 380 deletions
--- a/src/dream/dream_generator.py
+++ b/src/dream/dream_generator.py
@@ -0,0 +1,198 @@
+import random
+from typing import List, Optional
+
+from src.common.logger import get_logger
+from src.config.config import model_config
+from src.chat.utils.prompt_builder import Prompt
+from src.llm_models.payload_content.message import RoleType, Message
+from src.llm_models.utils_model import LLMRequest
+
+logger = get_logger("dream_generator")
+
+# 初始化 utils 模型用于生成梦境总结
+_dream_summary_model: Optional[LLMRequest] = None
+
+# 梦境风格列表（21种）
+DREAM_STYLES = [
+    "保持诗意和想象力，自由编写",
+    "诗意朦胧，如薄雾笼罩的清晨",
+    "奇幻冒险，充满未知与探索",
+    "温暖怀旧，带着时光的痕迹",
+    "神秘悬疑，暗藏深意",
+    "浪漫唯美，如诗如画",
+    "科幻未来，科技与想象交织",
+    "自然清新，如山林间的微风",
+    "深沉哲思，引人深思",
+    "轻松幽默，充满趣味",
+    "悲伤忧郁，带着淡淡哀愁",
+    "激昂热烈，充满活力",
+    "宁静平和，如湖面般平静",
+    "荒诞离奇，打破常规",
+    "细腻温柔，如春风拂面",
+    "壮阔宏大，气势磅礴",
+    "简约纯粹，返璞归真",
+    "复杂多变，层次丰富",
+    "梦幻迷离，虚实难辨",
+    "现实写意，贴近生活",
+    "抽象概念，超越具象",
+]
+
+
+def get_random_dream_styles(count: int = 2) -> List[str]:
+    """从梦境风格列表中随机选择指定数量的风格"""
+    return random.sample(DREAM_STYLES, min(count, len(DREAM_STYLES)))
+
+
+def get_dream_summary_model() -> LLMRequest:
+    """获取用于生成梦境总结的 utils 模型实例"""
+    global _dream_summary_model
+    if _dream_summary_model is None:
+        _dream_summary_model = LLMRequest(
+            model_set=model_config.model_task_config.utils,
+            request_type="dream.summary",
+        )
+    return _dream_summary_model
+
+
+def init_dream_summary_prompt() -> None:
+    """初始化梦境总结的提示词"""
+    Prompt(
+        """
+你刚刚完成了一次对聊天记录的记忆整理工作。以下是整理过程的摘要：
+整理过程：
+{conversation_text}
+
+请将这次整理涉及的相关信息改写为一个富有诗意和想象力的"梦境"，请你仅使用具体的记忆的内容，而不是整理过程编写。
+要求：
+1. 使用第一人称视角
+2. 叙述直白，不要复杂修辞，口语化
+3. 长度控制在200-800字
+4. 用中文输出
+梦境风格：
+{dream_styles}
+请直接输出梦境内容，不要添加其他说明：
+""",
+        name="dream_summary_prompt",
+    )
+
+
+async def generate_dream_summary(
+    chat_id: str,
+    conversation_messages: List[Message],
+    total_iterations: int,
+    time_cost: float,
+) -> None:
+    """生成梦境总结并输出到日志"""
+    try:
+        import json
+        from src.chat.utils.prompt_builder import global_prompt_manager
+        
+        # 第一步：建立工具调用结果映射 (call_id -> result)
+        tool_results_map: dict[str, str] = {}
+        for msg in conversation_messages:
+            if msg.role == RoleType.Tool and msg.tool_call_id:
+                content = ""
+                if msg.content:
+                    if isinstance(msg.content, list) and msg.content:
+                        content = msg.content[0].text if hasattr(msg.content[0], "text") else str(msg.content[0])
+                    else:
+                        content = str(msg.content)
+                tool_results_map[msg.tool_call_id] = content
+        
+        # 第二步：详细记录所有工具调用操作和结果到日志
+        tool_call_count = 0
+        logger.info(f"[dream][工具调用详情] 开始记录 chat_id={chat_id} 的所有工具调用操作：")
+        
+        for msg in conversation_messages:
+            if msg.role == RoleType.Assistant and msg.tool_calls:
+                tool_call_count += 1
+                # 提取思考内容
+                thought_content = ""
+                if msg.content:
+                    if isinstance(msg.content, list) and msg.content:
+                        thought_content = msg.content[0].text if hasattr(msg.content[0], "text") else str(msg.content[0])
+                    else:
+                        thought_content = str(msg.content)
+                
+                logger.info(f"[dream][工具调用详情] === 第 {tool_call_count} 组工具调用 ===")
+                if thought_content:
+                    logger.info(f"[dream][工具调用详情] 思考内容：{thought_content[:500]}{'...' if len(thought_content) > 500 else ''}")
+                
+                # 记录每个工具调用的详细信息
+                for idx, tool_call in enumerate(msg.tool_calls, 1):
+                    tool_name = tool_call.func_name
+                    tool_args = tool_call.args or {}
+                    tool_call_id = tool_call.call_id
+                    tool_result = tool_results_map.get(tool_call_id, "未找到执行结果")
+                    
+                    # 格式化参数
+                    try:
+                        args_str = json.dumps(tool_args, ensure_ascii=False, indent=2) if tool_args else "无参数"
+                    except Exception:
+                        args_str = str(tool_args)
+                    
+                    logger.info(f"[dream][工具调用详情] --- 工具 {idx}: {tool_name} ---")
+                    logger.info(f"[dream][工具调用详情] 调用参数：\n{args_str}")
+                    logger.info(f"[dream][工具调用详情] 执行结果：\n{tool_result}")
+                    logger.info(f"[dream][工具调用详情] {'-' * 60}")
+        
+        logger.info(f"[dream][工具调用详情] 共记录了 {tool_call_count} 组工具调用操作")
+        
+        # 第三步：构建对话历史摘要（用于生成梦境）
+        conversation_summary = []
+        for msg in conversation_messages:
+            role = msg.role.value if hasattr(msg.role, "value") else str(msg.role)
+            content = ""
+            if msg.content:
+                content = msg.content[0].text if isinstance(msg.content, list) and msg.content else str(msg.content)
+            
+            if role == "user" and "轮次信息" in content:
+                # 跳过轮次信息消息
+                continue
+            
+            if role == "assistant":
+                # 只保留思考内容，简化工具调用信息
+                if content:
+                    # 截取前500字符，避免过长
+                    content_preview = content[:500] + ("..." if len(content) > 500 else "")
+                    conversation_summary.append(f"[{role}] {content_preview}")
+            elif role == "tool":
+                # 工具结果，只保留关键信息
+                if content:
+                    # 截取前300字符
+                    content_preview = content[:300] + ("..." if len(content) > 300 else "")
+                    conversation_summary.append(f"[工具执行] {content_preview}")
+        
+        conversation_text = "\n".join(conversation_summary[-20:])  # 只保留最后20条消息
+        
+        # 随机选择2个梦境风格
+        selected_styles = get_random_dream_styles(2)
+        dream_styles_text = "\n".join([f"{i+1}. {style}" for i, style in enumerate(selected_styles)])
+        
+        # 使用 Prompt 管理器格式化梦境生成 prompt
+        dream_prompt = await global_prompt_manager.format_prompt(
+            "dream_summary_prompt",
+            chat_id=chat_id,
+            total_iterations=total_iterations,
+            time_cost=time_cost,
+            conversation_text=conversation_text,
+            dream_styles=dream_styles_text,
+        )
+
+        # 调用 utils 模型生成梦境
+        summary_model = get_dream_summary_model()
+        dream_content, (reasoning, model_name, _) = await summary_model.generate_response_async(
+            dream_prompt,
+            max_tokens=512,
+            temperature=0.8,
+        )
+        
+        if dream_content:
+            logger.info(f"[dream][梦境总结] 对 chat_id={chat_id} 的整理过程梦境：\n{dream_content}")
+        else:
+            logger.warning("[dream][梦境总结] 未能生成梦境总结")
+            
+    except Exception as e:
+        logger.error(f"[dream][梦境总结] 生成梦境总结失败: {e}", exc_info=True)
+
+init_dream_summary_prompt()