FinSight
经 AI Skill Hub 精选评估，FinSight 获评「推荐使用」。这款Agent工作流在功能完整性、社区活跃度和易用性方面表现出色，AI 评分 7.5 分，适合有一定技术背景的用户使用。
FinSight 工作流的设计遵循"最小配置，最大复用"原则：核心逻辑已经封装好，用户只需配置自己的 API Key 和业务参数即可快速上手。工作流内置错误处理和重试机制，在网络波动或 API 限速等情况下仍能稳定运行，适合作为生产环境的自动化基础设施。
在实际部署时，建议先在测试环境中运行 3-5 次，验证各个环节的输出结果符合预期，再部署到生产环境。AI Skill Hub 评分 7.5 分，是同类 Agent 工作流中的精选推荐。
Multi-agent会话式金融研究平台，提供专业智能体和实时流式响应
FinSight 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排，将复杂的多步骤任务拆解为清晰的自动化流程，实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成，适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
Multi-agent会话式金融研究平台，提供专业智能体和实时流式响应
FinSight 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排，将复杂的多步骤任务拆解为清晰的自动化流程，实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成，适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
- 可视化 Agent 工作流编排，无需编写复杂代码
- 支持多步骤自动化任务链，实现全流程无人值守
- 与外部 API、数据库和第三方服务无缝集成
- 内置错误处理与自动重试机制，保障稳定运行
- 提供可复用的自动化模板，快速在同类场景部署
- 自动化日常重复性工作，将精力集中于创造性任务
- 构建数据采集 → 处理 → 输出的完整自动化管线
- 实现跨平台、跨系统的数据流转和业务协同
# 方式一：pip 安装（推荐）
pip install finsight
# 方式二：虚拟环境安装（推荐生产环境）
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install finsight
# 方式三：从源码安装（获取最新功能）
git clone https://github.com/kkkano/FinSight
cd FinSight
pip install -e .
# 验证安装
python -c "import finsight; print('安装成功')"
- 访问 GitHub 仓库获取工作流文件
- 在对应平台（Dify / Flowise / Make 等）中找到「导入工作流」功能
- 上传工作流文件
- 按照提示配置必要的环境变量和 API Key
- 运行测试确认流程正常后投入使用
# 命令行使用
finsight --help
# 基本用法
finsight input_file -o output_file
# Python 代码中调用
import finsight
# 示例
result = finsight.process("input")
print(result)
# finsight 配置文件示例（config.yml） app: name: "finsight" debug: false log_level: "INFO" # 运行时指定配置文件 finsight --config config.yml # 或通过环境变量配置 export FINSIGHT_API_KEY="your-key" export FINSIGHT_OUTPUT_DIR="./output"
简介
<p align="center"> <img src="frontend/public/logo.svg" alt="FinSight AI Logo" width="80" height="80" /> </p>
FinSight AI
<p align="center"> <strong>Multi-Agent Financial Research Platform powered by LangGraph</strong> </p>
<p align="center"> <a href="./README.md">English</a> | <a href="./readme_cn.md">中文</a> | <a href="./docs/DOCS_INDEX.md">Docs Index</a> </p>
<p align="center"> 🌐 <strong>Live Demo:</strong> <a href="https://finsight-ai.chat">https://finsight-ai.chat</a> </p>
---
FinSight AI is a production-grade, multi-agent financial research system built on LangGraph. It unifies conversational AI analysis, a professional dashboard with 6 analytical tabs, autonomous task execution (Workbench), and proactive email alerts into one coherent platform.
7 Research Agents (autonomous, multi-tool) · 1 Synthesize Node (conflict detection + hallucination guard) · 5 Dashboard Scorers (per-tab AI cards) | Hybrid RAG (bge-m3) | Real-time ECharts | LLM-driven Smart Charts | Conflict detection across 8 agent pairs | Email subscription alerts
---
Overview Tab
Composite AI analysis with ScoreRing, Fear & Greed Gauge, Agent Coverage Matrix, Dimension Radar, Risk Metrics, Highlights, and Analyst Target Price.
💬 "Ask About This" (问这条)
A context-aware follow-up feature allowing users to ask AI about any specific news item, AI insight, or risk warning directly from the dashboard.
✨ Key Features
| Category | Highlights |
|---|---|
| **Multi-Agent Orchestration** | 7 specialized research agents (Price, News, Fundamental, Technical, Macro, Risk, DeepSearch) running in parallel execution groups |
| **LangGraph Pipeline** | Stateful LangGraph runtime for GPT-like chat, alerts, URL/article analysis, quick market answers, and explicit report generation. Ordinary chat uses the LLM conversation router before planning; only the report button enters the structured report template path. |
| **Professional Dashboard** | 6 analytical tabs (Overview, Financial, Technical, News, Research, Peers) with ECharts visualization |
| **AI-Powered Insights** | 5 Dashboard Scorers generate real-time AI analysis cards for each tab via single LLM call + deterministic fallback (1-3s each) |
| **Hybrid RAG Engine** | bge-m3 (1024-dim Dense + Sparse) with bge-reranker-v2-m3 cross-encoder reranking |
| **Smart Charts** | Dual-mode LLM-driven charts: <chart> (inline data) + <chart_ref> (real data reference) |
| **Conflict Detection** | Automatic cross-agent conflict analysis across 8 comparable dimension pairs |
| **Proactive Alerts** | 3 alert schedulers (Price, News, Risk) with email notification via SMTP |
| **Workbench** | Autonomous task execution, portfolio rebalancing with LLM enhancement, SSE streaming progress, report timeline, and quick analysis bar |
| **"Ask About This"** | Context-aware follow-up on any news, insight, or risk item via MiniChat integration |
| **ThinkingBubble** | Three-layer execution display: thinking bubble (typewriter effect) → agent summary cards → detailed timeline |
| **Morning Brief Pipeline** | One-click portfolio morning brief via LangGraph Pipeline with deterministic synthesis (zero LLM cost) |
| **Rebalance LLM Enhancement** | Agent-backed LLM priority refinement for rebalance suggestions with evidence snapshots |
| **Hallucination Defense** | Multi-layer scrubbing: regex pattern matching + evidence cross-validation on LLM outputs |
| **Conversational Price Alerts** | Chat-driven alert setup — say "alert me when AAPL drops below $180" → auto-extracted, persisted, and triggered by scheduler (Phase 1) |
| **Stock Screener** | Natural-language stock screening with multi-condition filters; capability_note boundary hints for CN/HK coverage (Phase 2) |
| **A-Share Market Data** | Northbound/Southbound capital flow, sector heat maps, concept board rankings for CN & HK markets (Phase 3) |
| **Strategy Backtesting** | SMA crossover, MACD, RSI strategies with T+1 settlement, cost/slippage modeling, and look-ahead bias prevention (Phase 4) |
---
🔑 API Keys (Required vs Optional)
| API Key | Required? | Purpose | If Not Configured |
|---|---|---|---|
OPENAI_COMPATIBLE_API_KEY | ✅ **Required** | Default OpenAI-compatible LLM endpoint (mimo-v2.5-pro service) | App won't function |
OPENAI_COMPATIBLE_API_BASE | ✅ **Required** | OpenAI-compatible base URL (https://token-plan-cn.xiaomimimo.com/v1 by default) | Uses the code default |
OPENAI_COMPATIBLE_MODEL | ✅ **Required** | Default model ID (mimo-v2.5-pro by default) | Uses the code default |
GEMINI_PROXY_API_KEY or OPENAI_API_KEY | Optional | Alternative LLM providers | OpenAI-compatible endpoint is used |
FMP_API_KEY | ⭐ Recommended | Financial data (earnings, ratios) | Falls back to yfinance |
FINNHUB_API_KEY | Optional | Real-time quotes, news | Falls back to other sources |
TAVILY_API_KEY | Optional | Web search | Falls back to DuckDuckGo |
FRED_API_KEY | Optional | Macro economic data | Limited macro features |
ALPHA_VANTAGE_API_KEY | Optional | Additional price data | Uses other price sources |
Minimum Setup: configureOPENAI_COMPATIBLE_API_KEYin.env.server. All other APIs have automatic fallbacks.
2. Install dependencies
pip install -r requirements.txt
🚀 Getting Started
🐳 Docker One-Click Deployment (Recommended)
```bash
Frontend Setup
```bash cd frontend pnpm install pnpm dev
2. Configure environment
cp .env.server.example .env.server
Edit .env.server with your real API keys (see "API Keys" section below)
1. Create virtual environment
python -m venv .venv
3. Configure environment
copy .env.server.example .env.server
Edit .env.server with your API keys:
Optional: PostgreSQL for RAG
```bash
Set environment variable to enable PostgreSQL backend
Optional: Email Alerts
```bash
LangChain / LangGraph APIs Used
| API | Usage |
|---|---|
langgraph.graph.MessagesState | Base state with add_messages reducer for conversation history |
langgraph.checkpoint.sqlite.SqliteSaver | Persistent conversation checkpoints (SQLite backend) |
langgraph.checkpoint.postgres.PostgresSaver | Optional PostgreSQL checkpoint backend |
langgraph.types.interrupt() | Human-in-the-loop pause at confirmation_gate |
langgraph.types.Command(resume=) | Resume execution after HITL approval |
langchain_core.messages.HumanMessage / SystemMessage / RemoveMessage | Message type construction |
langchain_core.messages.trim_messages | Context window management — trim old messages |
langfuse.decorators.langfuse_observe | Distributed tracing integration with Langfuse |
---
OPENAI_COMPATIBLE_API_KEY=sk-...
OPENAI_COMPATIBLE_API_BASE=https://token-plan-cn.xiaomimimo.com/v1
OPENAI_API_KEY=sk-...
GOOGLE_API_KEY=... (for Gemini)
FMP_API_KEY=... (Financial Modeling Prep)
FINNHUB_API_KEY=... (Finnhub)
TAVILY_API_KEY=... (Tavily Search)
FRED_API_KEY=... (FRED Economic Data)
🔄 LangGraph Pipeline (Conversational Runtime)
The FinSight chat runtime is a LangGraph stateful graph. The current path is prepare_context -> chat_respond -> understand_request: chat_respond only short-circuits pure social turns, while ordinary chat, follow-ups, non-financial boundaries, URL/page requests, alerts, market questions, and report requests go through the LLM conversation router inside understand_request.
The router decides whether the turn should be answered directly, clarified, turned into an alert, or planned as research. Research turns produce understanding, tasks[], blocked_tasks[], and a compatibility projection for the existing policy/planner/executor boundary. URL/page/article work is exposed as the planner/agent tool fetch_url_content; the request-understanding layer does not pre-fetch URLs.
understand_request also writes a structured ReplyContract so downstream nodes do not infer UX intent from raw keywords again. The current lanes are:
| Lane | Trigger | Output rule |
|---|---|---|
chat_answer | Default for ordinary explanations, follow-ups, corrections, and "no news / no links / direct answer" turns | Natural conversational answer; no forced news lookup and no report scaffold |
source_grounded_answer | Explicit news, links, URL/article reading, real-time quote, citation, or data-evidence request | Use source-capable tools; cite usable URLs or disclose that a usable source was unavailable |
report_generation | Report button, output_mode=investment_report, or explicit "generate report / research report" request | Use report structure and report citation policy |
Evidence is split from tool diagnostics. EvidenceItem/evidence_pool is reserved for usable source material. Tool failures such as 403, rejected, empty, timeout, or other failed outputs are recorded as ToolError rows in artifacts.tool_diagnostics and must not be rendered as news, sources, or conclusions.
The implementation and acceptance spec are tracked in docs/plans/2026-05-03_request_understanding_task_graph_spec.md. The current full chat UX acceptance run is docs/qa/chat-router-100-final100-current-state.md with JSON next to it: 100/100 PASS, including 95 hard red-line cases across context continuity, session isolation, report follow-ups, URL/news/quote source grounding, no-news correction, and tool-error evidence boundaries. The older 40-query run remains as legacy regression evidence in docs/qa/chat-ux-40-query-final40-post-context-binding.md.
Dashboard Scorers are served by /api/dashboard/insights and are not graph nodes in the chat pipeline.
Conversation UX is now split deliberately: the frontend keeps the active browser runtime in localStorage, while /api/conversations owns backend thread lifecycle and a lightweight server snapshot store for messages, title, pinned, and archive metadata. Creating, switching, renaming, and deleting conversations go through the backend API; deletion clears session context, report/citation index rows, thread RAG memory/working-set collections, and matching RAG observability runs. Stream stop uses AbortController plus backend cancellation events and executor/agent cancellation tokens, preserving partial answers instead of treating cancellation as an error.
Memory is scoped before routing. Durable user preferences and historical focus are loaded for personalization, but only current_thread_focus and current_report from the active thread_id can bind deictic follow-ups such as "that report" or "the third point". Legacy user-level last_report, last_focus, and recent_focuses are kept as historical memory and are not exposed to the conversation router as current referents.
Runtime preferences are passed with chat options as agent_preferences. timeoutSeconds=0 means system default; positive values are clamped to 30-1200 seconds and applied to chat, planner, synthesis, and both sync/stream graph execution budgets.
Parse input, load memory"] INIT --> MEMSCOPE["memory_scope
user profile + current thread only"] MEMSCOPE --> RESET["② reset_turn_state
Clear ephemeral fields + trace runtime"] RESET --> PREPARE["③ prepare_context
Bound history, summarize, normalize UI hints"] PREPARE --> CHATRESP["④ chat_respond
Pure social only"] CHATRESP -->|"pure greeting / thanks / bye"| END_CHAT((End)) CHATRESP -->|"all other turns"| UNDERSTAND{"⑤ understand_request
LLM router + tasks + trace"} UNDERSTAND -->|"direct / clarify / out_of_scope"| END_CHAT UNDERSTAND -->|"alert"| ALERT_EX["⑥ alert_extractor
Extract alert params"] ALERT_EX -->|"valid"| ALERT_ACT["⑥b alert_action
Save & schedule"] ALERT_EX -->|"invalid"| END_ALERT_INVALID((End)) ALERT_ACT --> END_ALERT((End)) UNDERSTAND -->|"research"| POLICY["⑦ policy_gate
Capability scoring + task tool union"] POLICY --> PLAN["⑧ planner
LLM Planner or Stub Fallback"] PLAN --> CONFIRM{"⑨ confirmation_gate
HITL approval?"} CONFIRM -->|"cancel"| END_CANCEL((End)) CONFIRM -->|"adjust"| PLAN CONFIRM -->|"confirm"| EXEC["⑩ execute_plan
Tools, agents, URL fetch"] EXEC --> SYNTH["⑪ synthesize
Merge outputs + conflict/evidence checks"] SYNTH --> RENDER["⑫ render
Chat or report output"] RENDER --> END((End)) subgraph "Execution Engine (⑩)" direction LR EG1["Group 1
price · news"] --> EG2["Group 2
fundamental · technical"] EG2 --> EG3["Group 3
macro · risk · deep_search"] end EXEC -.-> EG1 style RESET fill:#a855f7,color:#fff style SYNTH fill:#ff9800,color:#000 style POLICY fill:#2196f3,color:#fff
report_builder and hallucination/evidence checks are implementation helpers inside the graph runtime, not separate graph nodes in backend/graph/runner.py.
🔍 RAG Engine — Hybrid Search Pipeline
FinSight uses a production-grade hybrid retrieval pipeline replacing the legacy SHA1 hash-based pseudo-embeddings.
SKIP / SECONDARY / PRIMARY"] ROUTER -->|PRIMARY| EMBED["bge-m3 Encode
1024-dim Dense + Sparse lexical"] ROUTER -->|SKIP| LIVE["Live Tools Only"] EMBED --> DENSE["Dense Search
Cosine similarity"] EMBED --> SPARSE["Sparse Search
Lexical weight matching"] DENSE --> RRF["RRF Fusion
+ Scope Boost"] SPARSE --> RRF RRF --> RERANK["Cross-Encoder Rerank
bge-reranker-v2-m3
Top-30 → Top-8"] RERANK --> OUTPUT["rag_context
Injected into synthesize prompt"] style EMBED fill:#4caf50,color:#fff style RERANK fill:#ff9800,color:#000
Key Components
| Component | File | Model / Algorithm |
|---|---|---|
| **Embedder** | rag/embedder.py | BAAI/bge-m3 — 1024-dim Dense + Sparse (lexical weights) |
| **Hybrid Search** | rag/hybrid_service.py | RRF fusion with scope boosting: persistent +0.15, medium_ttl +0.05 |
| **Reranker** | rag/reranker.py | BAAI/bge-reranker-v2-m3 Cross-Encoder, Top-30 → Top-8 |
| **Router** | rag/rag_router.py | Rule-based: SKIP (realtime quotes) / PRIMARY (historical) / PARALLEL (deep research) |
| **Chunker** | rag/chunker.py | Per-doc-type strategy: news (no split) / filings (1000/200) / transcripts (800/100) |
| **Store** | rag/hybrid_service.py | In-Memory or PostgreSQL (pgvector VECTOR(1024) + tsvector) |
Email Pipeline
Scheduler → Rule Engine → Alert Created →
HTML Template (Jinja2) → SMTP Send →
Delivery Tracking (transient vs permanent errors) →
Auto-disable after 3 permanent failuresManual Setup (Alternative)
Prerequisites
- Python 3.11+
- Node.js 18+ with pnpm
- At least one LLM API key (OpenAI / Gemini / DeepSeek)
Backend Setup
```bash
FinSight是一个开源的AI工作流平台，提供多种智能体和实时流式响应功能，适合金融研究和投资组合监控，但需要进一步优化和完善
- 构建多智能体协作系统的 Agent 开发者
- 构建企业知识库 / RAG 检索应用的团队
- 生产部署优先使用 Docker Compose 隔离依赖，并挂载 volume 持久化数据
- 分块大小建议 256-512 tokens，向量库优选 pgvector 或 Qdrant
- Agent 任务先做 dry-run 验证工具调用链，再开启自主执行
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
- 容器内无法访问宿主机 localhost — 使用 host.docker.internal
- embedding 模型与查询模型不一致导致检索失效
- Python 依赖冲突：建议用 venv / uv 隔离环境
- Docker：FinSight 提供官方镜像，docker compose up 一键启动
- 云端托管：可放在 Vercel / Railway / Fly.io 等 PaaS 平台
- 可视化 Agent 工作流编排，无需编写复杂代码
- 支持多步骤自动化任务链，实现全流程无人值守
- 与外部 API、数据库和第三方服务无缝集成
- 内置错误处理与自动重试机制，保障稳定运行
- 提供可复用的自动化模板，快速在同类场景部署
- 构建多智能体协作系统的 Agent 开发者
- 构建企业知识库 / RAG 检索应用的团队
- 生产部署优先使用 Docker Compose 隔离依赖，并挂载 volume 持久化数据
- 分块大小建议 256-512 tokens，向量库优选 pgvector 或 Qdrant
- Agent 任务先做 dry-run 验证工具调用链，再开启自主执行
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
- 容器内无法访问宿主机 localhost — 使用 host.docker.internal
- embedding 模型与查询模型不一致导致检索失效
- Python 依赖冲突：建议用 venv / uv 隔离环境
- 自动化日常重复性工作，将精力集中于创造性任务
- 构建数据采集 → 处理 → 输出的完整自动化管线
- 实现跨平台、跨系统的数据流转和业务协同
- +MIT 协议，可免费商用
- +大幅减少重复性人工操作
- +可视化流程，清晰直观
- +可扩展性强，支持复杂场景
- −初始配置和调试需投入一定时间
- −强依赖外部服务的稳定性
- −复杂场景需具备一定技术基础
AI Skill Hub 为第三方内容聚合平台，本页面信息基于公开数据整理，不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后，再部署至生产环境，并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一，可自由商用、修改、分发，仅需保留版权声明。
AI Skill Hub 点评：FinSight 的核心功能完整，质量良好。对于自动化工程师和运维人员来说，这是一个值得纳入个人工具库的选择。建议先在非生产环境试用，再逐步推广。
| 原始名称 | FinSight |
| 原始描述 | 开源AI工作流：Multi-agent会话式金融研究平台。提供专业智能体（价格分析、新闻追踪、宏观经济、深度研究），支持实时流式响应、自动生成8章节投资报告和投资组合监控。采用。⭐31 · Python |
| Topics | workflowai-agentsfinancial-analysis |
| GitHub | https://github.com/kkkano/FinSight |
| License | MIT |
| 语言 | Python |
收录时间：2026-05-22 · 更新时间：2026-05-23 · License：MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。