kunthawat/microfish
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: translate Chinese prompts to English in simulation config and profile generators

Browse Source 
- simulation_config_generator.py: translate all LLM prompts and system messages
- oasis_profile_generator.py: translate profile generation prompts

Ensures get_language_instruction() controls output language instead of
being overridden by Chinese prompt context.
This commit is contained in:
Kunthawat Greethong
2026-06-17 15:52:38 +07:00
parent bf14c56944
commit 5fcce79361
2 changed files with 30 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
backend/app/services/oasis_profile_generator.py
View File

@@ -671,7 +671,7 @@ class OasisProfileGenerator:
def _get_system_prompt(self, is_individual: bool) -> str:
"""获取系统提示词"""
base_prompt = "你是社交媒体用户画像生成专家。生成详细、真实的人设用于舆论模拟,最大程度还原已有现实情况。必须返回有效的JSON格式，所有字符串值不能包含未转义的换行符。"
base_prompt = "You are a social media user persona generation expert. Generate detailed, realistic personas for opinion simulation, maximizing fidelity to real-world situations. Must return valid JSON format, all string values must not contain unescaped newlines."
return f"{base_prompt}\n\n{get_language_instruction()}"
def _build_individual_persona_prompt(
@@ -687,7 +687,7 @@ class OasisProfileGenerator:
attrs_str = json.dumps(entity_attributes, ensure_ascii=False) if entity_attributes else "无"
context_str = context[:3000] if context else "无额外上下文"
return f"""为实体生成详细的社交媒体用户人设,最大程度还原已有现实情况。
return f"""Generate detailed social media user persona for the entity, maximizing fidelity to real-world situations.
实体名称: {entity_name}
实体类型: {entity_type}
@@ -736,7 +736,7 @@ class OasisProfileGenerator:
attrs_str = json.dumps(entity_attributes, ensure_ascii=False) if entity_attributes else "无"
context_str = context[:3000] if context else "无额外上下文"
return f"""为机构/群体实体生成详细的社交媒体账号设定,最大程度还原已有现实情况。
return f"""Generate detailed social media account settings for organization/group entity, maximizing fidelity to real-world situations.
实体名称: {entity_name}
实体类型: {entity_type}
@@ -769,7 +769,7 @@ class OasisProfileGenerator:
- persona必须是一段连贯的文字描述，不要使用换行符
- {get_language_instruction()} (gender字段必须用英文"other")
- age必须是整数30，gender必须是字符串"other"
- 机构账号发言要符合其身份定位"""
- Organization account posts must match its identity positioning"""
def _generate_profile_rule_based(
self,
@@ -1076,11 +1076,11 @@ class OasisProfileGenerator:
- name: 用户真实姓名
- username: 系统中的用户名
- user_char: 详细人设描述（注入到LLM系统提示中，指导Agent行为）
- description: 简短的公开简介（显示在用户资料页面）
- description: Short public bio (displayed on user profile page)
user_char vs description 区别：
user_char vs description difference:
- user_char: 内部使用，LLM系统提示，决定Agent如何思考和行动
- description: 外部显示，其他用户可见的简介
- description: Externally displayed, visible to other users
"""
import csv
@@ -1112,7 +1112,7 @@ class OasisProfileGenerator:
profile.name, # name: 真实姓名
profile.user_name, # username: 用户名
user_char, # user_char: 完整人设（内部LLM使用）
description # description: 简短简介（外部显示）
description # description: Short bio (externally displayed)
]
writer.writerow(row)

44
backend/app/services/simulation_config_generator.py
View File

@@ -540,15 +540,15 @@ class SimulationConfigGenerator:
# 计算最大允许值（80%的agent数）
max_agents_allowed = max(1, int(num_entities * 0.9))
prompt = f"""基于以下模拟需求，生成时间模拟配置。
prompt = f"""Based on the following simulation requirements, generate time simulation config.
{context_truncated}
## 任务
请生成时间配置JSON。
Please generate time config JSON.
### 基本原则（仅供参考，需根据具体事件和参与群体灵活调整）：
- 请根据模拟场景推断目标用户群体所在时区和作息习惯，以下为东八区(UTC+8)的参考示例
- Please infer the target user group's timezone and daily routine based on the simulation scenario. Below is a reference example for UTC+8
- 凌晨0-5点几乎无人活动（活跃度系数0.05）
- 早上6-8点逐渐活跃（活跃度系数0.4）
- 工作时间9-18点中等活跃（活跃度系数0.7）
@@ -571,7 +571,7 @@ class SimulationConfigGenerator:
"off_peak_hours": [0, 1, 2, 3, 4, 5],
"morning_hours": [6, 7, 8],
"work_hours": [9, 10, 11, 12, 13, 14, 15, 16, 17, 18],
"reasoning": "针对该事件的时间配置说明"
"reasoning": "Explanation of time config for this event"
}}
字段说明：
@@ -583,15 +583,15 @@ class SimulationConfigGenerator:
- off_peak_hours (int数组): 低谷时段，通常深夜凌晨
- morning_hours (int数组): 早间时段
- work_hours (int数组): 工作时段
- reasoning (string): 简要说明为什么这样配置"""
- reasoning (string): Brief explanation of why this config was chosen"""
system_prompt = "你是社交媒体模拟专家。返回纯JSON格式，时间配置需符合模拟场景中目标用户群体的作息习惯。"
system_prompt = "You are a social media simulation expert. Return pure JSON format. Time config must match the target user group's daily routine in the simulation scenario."
system_prompt = f"{system_prompt}\n\n{get_language_instruction()}"
try:
return self._call_llm_with_retry(prompt, system_prompt)
except Exception as e:
logger.warning(f"时间配置LLM生成失败: {e}, 使用默认配置")
logger.warning(f"Time config LLM generation failed: {e}, using default config")
return self._get_default_time_config(num_entities)
def _get_default_time_config(self, num_entities: int) -> Dict[str, Any]:
@@ -605,7 +605,7 @@ class SimulationConfigGenerator:
"off_peak_hours": [0, 1, 2, 3, 4, 5],
"morning_hours": [6, 7, 8],
"work_hours": [9, 10, 11, 12, 13, 14, 15, 16, 17, 18],
"reasoning": "使用默认中国人作息配置（每轮1小时）"
"reasoning": "Using default UTC+8 daily routine config (1 hour per round)"
}
def _parse_time_config(self, result: Dict[str, Any], num_entities: int) -> TimeSimulationConfig:
@@ -673,9 +673,9 @@ class SimulationConfigGenerator:
# 使用配置的上下文截断长度
context_truncated = context[:self.EVENT_CONFIG_CONTEXT_LENGTH]
prompt = f"""基于以下模拟需求，生成事件配置。
prompt = f"""Based on the following simulation requirements, generate event config.
模拟需求: {simulation_requirement}
Simulation requirements: {simulation_requirement}
{context_truncated}
@@ -683,7 +683,7 @@ class SimulationConfigGenerator:
{type_info}
## 任务
请生成事件配置JSON：
Please generate event config JSON:
- 提取热点话题关键词
- 描述舆论发展方向
- 设计初始帖子内容，**每个帖子必须指定 poster_type（发布者类型）**
@@ -691,7 +691,7 @@ class SimulationConfigGenerator:
**重要**: poster_type 必须从上面的"可用实体类型"中选择，这样初始帖子才能分配给合适的 Agent 发布。
例如：官方声明应由 Official/University 类型发布，新闻由 MediaOutlet 发布，学生观点由 Student 发布。
返回JSON格式（不要markdown）：
Return in JSON format (no markdown):
{{
"hot_topics": ["关键词1", "关键词2", ...],
"narrative_direction": "<舆论发展方向描述>",
@@ -702,18 +702,18 @@ class SimulationConfigGenerator:
"reasoning": "<简要说明>"
}}"""
system_prompt = "你是舆论分析专家。返回纯JSON格式。注意 poster_type 必须精确匹配可用实体类型。"
system_prompt = "You are a public opinion analysis expert. Return pure JSON format. Note: poster_type must exactly match available entity types."
system_prompt = f"{system_prompt}\n\n{get_language_instruction()}\nIMPORTANT: The 'poster_type' field value MUST be in English PascalCase exactly matching the available entity types. Only 'content', 'narrative_direction', 'hot_topics' and 'reasoning' fields should use the specified language."
try:
return self._call_llm_with_retry(prompt, system_prompt)
except Exception as e:
logger.warning(f"事件配置LLM生成失败: {e}, 使用默认配置")
logger.warning(f"Event config LLM generation failed: {e}, using default config")
return {
"hot_topics": [],
"narrative_direction": "",
"initial_posts": [],
"reasoning": "使用默认配置"
"reasoning": "Using default config"
}
def _parse_event_config(self, result: Dict[str, Any]) -> EventConfig:
@@ -830,9 +830,9 @@ class SimulationConfigGenerator:
"summary": e.summary[:summary_len] if e.summary else ""
})
prompt = f"""基于以下信息，为每个实体生成社交媒体活动配置。
prompt = f"""Based on the following information, generate social media activity config for each entity.
模拟需求: {simulation_requirement}
Simulation requirements: {simulation_requirement}
## 实体列表
```json
@@ -840,14 +840,14 @@ class SimulationConfigGenerator:
```
## 任务
为每个实体生成活动配置，注意：
- **时间符合目标用户群体作息**：以下为参考（东八区），请根据模拟场景调整
Generate activity config for each entity, noting:
- **Time must match target user group routine**: Below is a reference (UTC+8), please adjust based on simulation scenario
- **官方机构**（University/GovernmentAgency）：活跃度低(0.1-0.3)，工作时间(9-17)活动，响应慢(60-240分钟)，影响力高(2.5-3.0)
- **媒体**（MediaOutlet）：活跃度中(0.4-0.6)，全天活动(8-23)，响应快(5-30分钟)，影响力高(2.0-2.5)
- **个人**（Student/Person/Alumni）：活跃度高(0.6-0.9)，主要晚间活动(18-23)，响应快(1-15分钟)，影响力低(0.8-1.2)
- **公众人物/专家**：活跃度中(0.4-0.6)，影响力中高(1.5-2.0)
返回JSON格式（不要markdown）：
Return in JSON format (no markdown):
{{
"agent_configs": [
{{
@@ -866,14 +866,14 @@ class SimulationConfigGenerator:
]
}}"""
system_prompt = "你是社交媒体行为分析专家。返回纯JSON，配置需符合模拟场景中目标用户群体的作息习惯。"
system_prompt = "You are a social media behavior analysis expert. Return pure JSON. Config must match the target user group's daily routine in the simulation scenario."
system_prompt = f"{system_prompt}\n\n{get_language_instruction()}\nIMPORTANT: The 'stance' field value MUST be one of the English strings: 'supportive', 'opposing', 'neutral', 'observer'. All JSON field names and numeric values must remain unchanged. Only natural language text fields should use the specified language."
try:
result = self._call_llm_with_retry(prompt, system_prompt)
llm_configs = {cfg["agent_id"]: cfg for cfg in result.get("agent_configs", [])}
except Exception as e:
logger.warning(f"Agent配置批次LLM生成失败: {e}, 使用规则生成")
logger.warning(f"Agent config batch LLM generation failed: {e}, using rule-based generation")
llm_configs = {}
# 构建AgentActivityConfig对象