A universal Chinese public opinion monitoring system. Supports any company, brand, or person as the monitoring target. Features four-level alert system, SnowNLP local sentiment classification (zero token), LLM precision judgment for uncertain items, semantic deduplication, heat scoring, and trend tracking.
All target-specific data (keywords, recipients, etc.) is driven by a JSON config file — no hardcoding.
When the user triggers this skill for the first time, collect the following information:
sentiment_data/ under workspace)
Generate a monitor_config.json in the user's workspace using this template:
{
"target_name": "目标名称",
"target_aliases": ["别名1", "别名2"],
"keywords": {
"core": ["核心关键词1", "核心关键词2"],
"products": [],
"tech": [],
"people": [],
"variants": []
},
"email_to": "recipient@example.com",
"platforms": [],
"data_dir": "sentiment_data",
"created_at": "2026-04-18T00:00:00"
}
Use the generate_config() function from scripts/sentiment_engine.py to create this programmatically, or write it directly.
pip install snownlp agentmail-sdk
SnowNLP (~7.7MB) provides local Chinese sentiment classification at 280 items/sec with zero token cost. AgentMail SDK is needed only if email delivery is required.
Use automation_update tool to create scheduled report tasks. Each task's prompt should follow the Report Workflow below.
Each scheduled report execution follows this workflow. Embed this in the automation prompt:
import sys
sys.path.insert(0, r"<SKILL_SCRIPTS_DIR>")
# <SKILL_SCRIPTS_DIR> is the absolute path to this skill's scripts/ directory
# On Windows: C:\Users\<USER>\.workbuddy\skills\sentiment-monitor\scripts
from sentiment_engine import *
from report_generator import make_html_report, make_text_report
cfg = load_config("<WORKSPACE>/monitor_config.json")
history = load_json(get_history_file(cfg), {"items": []})
Read keywords from config:
keywords = get_all_keywords(cfg)
kw_counts = get_keyword_count(cfg)
Search all keywords across web, news, and social media platforms. Structure each result as:
{
"title": "标题",
"summary": "摘要",
"source": "来源平台/媒体",
"url": "链接",
"time": "发布时间",
}
all_items = deduplicate_items(raw_items)
Apply the three-layer classification strategy:
Layer 1 — SnowNLP Local Classification (zero token)
results, uncertain_indices = batch_snow_classify(all_items)
score > 0.70 → positive (confident, no LLM needed)
score < 0.30 → negative (confident, no LLM needed)
0.30 ~ 0.70 → uncertain, needs LLM
Layer 2 — LLM Precision Judgment (only for uncertain items)
Send only items at uncertain_indices to LLM for classification. This typically covers 20-40% of total items, saving 60-80% of tokens.
Layer 3 — Alert Keyword Check (zero token, pure keyword matching)
For all items classified as negative, run assess_alert_level() to determine alert level:
for i, item in enumerate(all_items):
item["heat_score"] = calc_heat_score(item.get("source"), i // len(all_items) * 10 + 1, item.get("time"))
item["sentiment"] = results[i]["sentiment"]
item["alert_level"] = assess_alert_level(item, all_items)
Compare with history.json to find new items. Get trend data:
trend = get_trend_summary(cfg)
since_label = time_ago_str(history.get("last_report_time"))
kw_summary = f"{kw_counts['core']}核心 + {kw_counts['total'] - kw_counts['core']}关联"
highest = get_highest_alert([i.get("alert_level") for i in new_items])
html = make_html_report(
target_name=cfg["target_name"],
pos_items=positive_items,
neg_items=negative_items,
neu_items=neutral_items,
highest_alert=highest,
trend_data=trend,
summary_html=summary_html,
run_label=get_run_label(),
since_label=since_label,
keyword_summary=kw_summary,
snownlp_count=snownlp_classified_count,
llm_count=llm_classified_count,
)
text = make_text_report(
target_name=cfg["target_name"],
# ... same args ...
summary_text=summary_text,
)
from report_sender import send_sentiment_report
send_sentiment_report(
target_name=cfg["target_name"],
run_label=get_run_label(),
html_content=html,
text_content=text,
to_email=cfg["email_to"],
)
record_trend(cfg, pos_count, neg_count, neu_count, highest, get_run_label())
# Also update history.json with new items and last_report_time
Copy and customize this template for each scheduled report:
执行 {target_name} 舆情监控任务（{早报/午报/晚报}），按以下步骤执行：
## 0. 导入引擎
import sys; sys.path.insert(0, r"C:\Users\{USER}\.workbuddy\skills\sentiment-monitor\scripts")
from sentiment_engine import *
from report_generator import make_html_report, make_text_report
## 1. 加载配置
读取 {WORKSPACE}/monitor_config.json 和历史数据
## 2. 全网搜索（{keyword_count}个关键词）
核心：{列出核心关键词}
产品：{列出产品关键词}
技术：{列出技术关键词}
关联：{列出关联关键词}
变体：{列出变体关键词}
## 3. 智能去重
使用 deduplicate_items() 去重
## 4. 情感分类（省token策略）
第一步：SnowNLP 本地初筛（零token）
第二步：仅不确定条目调 LLM 精判
第三步：预警关键词兜底（零token）
## 5. 增量对比 + 热度评分 + 预警评估
## 6. 生成 HTML+纯文本 报告
## 7. 发送邮件至 {email_to}
## 8. 更新数据文件（history.json + trends.json）
from sentiment_engine import snownlp_classify
sentiment, confidence, needs_llm = snownlp_classify("文本内容")
python sentiment_engine.py <config.json> keywords # 列出关键词
python sentiment_engine.py <config.json> alerts # 查看预警配置
python sentiment_engine.py <config.json> trends # 查看趋势
python sentiment_engine.py <config.json> classify <文本> # 测试分类
python sentiment_engine.py <config.json> benchmark # 基准测试
For detailed architecture and design decisions, read references/architecture.md.
| File | Purpose |
|------|---------|
| scripts/sentiment_engine.py | Core engine: config, dedup, SnowNLP classification, heat score, alert level, trends |
| scripts/report_generator.py | HTML + text report generation |
| scripts/report_sender.py | Email delivery via AgentMail |
| Package | Size | Purpose | Required |
|---------|------|---------|----------|
| snownlp | ~7.7MB | Local Chinese sentiment classification | Yes |
| agentmail-sdk | ~1MB | Email delivery | Only if sending reports |
共 1 个版本