Mine Xiaohongshu comments. Discover opportunities. Generate reports with evidence.
English — RedNote Insight analyzes Xiaohongshu (RedNote) product reviews to generate two parallel AI reports from a single search: a Product Selection Report for e-commerce sellers and a Content Strategy Plan for bloggers. Every insight is backed by cited review evidence — no hallucinated numbers.
中文 — 输入一个品类，同时生成两份 AI 报告：面向电商卖家的选品分析和面向内容博主的选题方案。所有结论绑定评论原文证据，杜绝 LLM 编造数据。
| Report · 报告 | Audience · 受众 | Includes · 包含 |
|---|---|---|
| 📊 Product Selection · 选品报告 | E-commerce sellers · 电商卖家 | Pain points, demand signals, brand distribution, evidence comments · 用户痛点、需求信号、品牌分布、证据评论 |
| 🎬 Content Strategy · 选题方案 | Content creators · 内容博主 | 3 viral topics + script outlines + cover design + publishing strategy · 3个爆款选题 + 脚本大纲 + 封面方案 + 发布策略 |
English — Don't know what to search? The sidebar "Inspiration Library" offers 189 curated search directions across 9 categories. Each entry is tagged (🛒 product / 🎬 content / 🛒+🎬 both) with a one-line prompt. Click to search instantly.
中文 — 不知道搜什么？左侧栏「灵感库」提供 9 个品类 × 21 条 = 189 条精选方向。每条标注适用标签（🛒选品 / 🎬选题 / 🛒+🎬双用），配一句话方向提示。点击直接搜，零等待。
┌──────────────────────────────────────────────────┐
│ Frontend · 前端 │
│ React 18 + Vite 5 · SSE streaming · 9 components │
├──────────────────────────────────────────────────┤
│ API · 接口 │
│ FastAPI async · 11 routes · dependency injection │
├──────────────────────────────────────────────────┤
│ Agent Pipeline · Agent 管道 │
│ Hybrid Retrieval → Comment Analysis → Aggregation│
│ ├→ InsightGenerator → 📊 Product Report │
│ └→ CreatorGenerator → 🎬 Content Plan │
├──────────────────────────────────────────────────┤
│ Data · 数据 │
│ PostgreSQL/pgvector · ChromaDB · Redis │
└──────────────────────────────────────────────────┘
Data Flow · 数据流: User input → Hybrid RAG (BGE-M3 + BM25 + RRF) → CrossEncoder Rerank → Comment Analysis + Demand Aggregation → Dual Agent parallel generation → SSE streaming
English — BGE-M3 vector search captures semantic similarity. BM25 + jieba ensures exact brand/model matching. RRF (K=60) fuses heterogeneous scores. CrossEncoder reranks the top results — 10× faster than LLM-as-Judge.
中文 — BGE-M3 向量检索捕捉中文语义相似性，BM25 + jieba 精确匹配品牌名和型号，RRF（K=60）融合异构分数，CrossEncoder 重排序比 LLM-as-Judge 快 10 倍。
English — One retrieval, one aggregation, two agents. InsightGenerator produces product selection reports with evidence citations; CreatorGenerator generates content plans with scripts and cover strategies. Both stream simultaneously via a single SSE connection.
中文 — 一次检索、一次聚合，双 Agent 并行：InsightGenerator 生成带证据引用的选品报告，CreatorGenerator 生成选题脚本和封面方案。两路 Token 经单一 SSE 连接同时推送。
English — ReportGuard validates numbers against actual review data. Business metrics (profit, weight, shipping cost) that aren't present in the review data are explicitly returned as null — the system refuses to hallucinate entry-barrier conclusions.
中文 — ReportGuard 按金额、比例、评分、重量和销量进行分语义校验。经营数据缺失时，相关字段显式返回 null，拒绝编造入场结论。
English — Rebuilt from vanilla JS to React 18 + Vite 5. Custom useSSE hook with request deduplication and AbortController. 9 independent components with Context + useReducer state management.
中文 — 从原生 JS 重构为 React 18 + Vite 5。自定义 useSSE Hook 实现请求去重和 AbortController。9 个独立组件，Context + useReducer 状态管理。
English — Docker + docker-compose for one-command deployment (API + PostgreSQL + Redis). GitHub Actions CI: lint → test → build. 140 automated tests including 21 trust-generation-specific tests.
中文 — Docker + docker-compose 一键部署（API + PostgreSQL + Redis）。GitHub Actions CI 流水线：lint → test → build。140 项自动化测试，含 21 项可信生成专项测试。
- Python 3.11+
- uv (auto-installed by run script)
- API key from SiliconFlow or DeepSeek
# Windows · Windows 系统
run.bat
# macOS / Linux
chmod +x run.sh && ./run.shgit clone https://github.com/Amazinghorseli/RedNote-Insight.git
cd RedNote-Insight
cp .env.example .env # Edit .env with your API key · 编辑 .env 填入 API Key
uv sync # Install dependencies · 安装依赖
uv run python generate_data.py # Generate demo data · 生成演示数据
uv run uvicorn api:app --host 0.0.0.0 --port 8000 --reloadThen open · 然后打开 http://localhost:8000 and http://localhost:8000/docs (Swagger).
docker-compose up -d| Layer · 层级 | Technology · 技术 | Purpose · 用途 |
|---|---|---|
| Frontend · 前端 | React 18, Vite 5, SSE | SPA with real-time streaming · 流式双报告 UI |
| Backend · 后端 | FastAPI, Pydantic v2 | Async API with dependency injection · 全异步 API |
| LLM · 大模型 | DeepSeek V4 | Report generation · 双报告生成 |
| Embedding · 向量 | BGE-M3 | Semantic search for Chinese text · 中文语义检索 |
| Retrieval · 检索 | BM25 + RRF + CrossEncoder | Hybrid retrieval pipeline · 混合检索引擎 |
| Vector DB · 向量库 | ChromaDB / pgvector | Comment storage & search · 评论存储与检索 |
| Task Queue · 任务 | Celery + Redis | Async ingestion & monitoring · 异步数据导入与监测 |
| Database · 数据库 | PostgreSQL | Business metrics & snapshots · 经营指标与快照 |
| Testing · 测试 | pytest, Vitest | 140 unit tests + E2E · 140 项单测 |
| Infra · 基础设施 | Docker, GitHub Actions | CI/CD pipeline · 持续集成 |
| Endpoint | Method | Description · 描述 |
|---|---|---|
/api/insight/stream |
POST | Generate dual reports via SSE · SSE 流式生成双报告 |
/api/insight |
POST | Generate reports (non-streaming) · 非流式生成报告 |
/api/qa |
POST | Q&A over report data · 报告数据问答 |
/api/trending |
GET | Trending topics & alerts · 趋势话题与告警 |
/api/inspiration |
GET | Inspiration library entries · 灵感库条目 |
/health |
GET | Health check · 健康检查 |
Full docs · 完整文档: http://localhost:8000/docs
RedNote-Insight/
├── frontend/ # React 18 + Vite 5 (9 components)
├── src/
│ ├── api/ # FastAPI routes (11 routes)
│ ├── agents/ # InsightGenerator, CreatorGenerator, comment/demand agents
│ ├── prompts/ # YAML-managed prompt templates
│ ├── data/ # Inspiration library (189 curated directions)
│ ├── domain/ # Pydantic models & schemas
│ ├── pipelines/ # Ingestion & monitoring pipelines
│ ├── repositories/ # Data access layer
│ ├── retrievers.py # Hybrid retrieval + RRF + reranker
│ ├── config.py # Settings via pydantic-settings
│ └── crawler.py # Data collection
├── tests/ # 140 unit tests
├── data/ # Demo data & ChromaDB
├── Dockerfile
├── docker-compose.yml
├── run.bat / run.sh # One-click launchers
└── pyproject.toml
- Hybrid RAG retrieval (BGE-M3 + BM25 + RRF + CrossEncoder)
- Dual Agent SSE streaming
- ReportGuard trusted generation
- React + Vite frontend migration
- Inspiration library (189 curated directions)
- Docker + CI/CD pipeline
- Upgrade to DeepSeek V4
- Public demo deployment
- Multi-platform review support (Douyin, Taobao)
- Real-time trend monitoring dashboard
MIT © 2026
Built with ❤️ for makers and sellers · 为创业者和卖家而生