PortfolioIQ: An Intelligent Investment Analysis System with Four-Agent Collaboration

This article introduces the open-source project PortfolioIQ, a multi-agent financial analysis workflow built on LangGraph. Through the collaborative division of labor among four agents—Researcher, Analyst, Critic, and Decision—and combining RAG retrieval and structured review mechanisms, it addresses the problems of information overload and decision complexity in the financial investment field, providing intelligent analysis support for individual investors and institutions.

In the field of financial investment, information overload and decision complexity are two major challenges faced by investors. A single agent struggles to balance depth and breadth, while multi-agent systems improve overall output quality through role division and collaboration. Financial investment analysis is an ideal scenario for multi-agent systems: it requires multi-source data integration, combination of quantitative and qualitative analysis, equal emphasis on risks and opportunities, and critical review to avoid bias.

PortfolioIQ is a four-agent workflow based on LangGraph, simulating the collaboration of a professional investment team. The tech stack includes: LangGraph (agent framework), Yahoo Finance + FAISS (data retrieval), FastAPI (API service), Docker + GCP Cloud Run (deployment), LangSmith (observability), and FAISS (vector storage). The project has a clear structure, including the app directory (main program, graph, state, agents, tools, prompts) and the tests directory.

The collaboration process is divided into four phases: 1. Researcher: Collects real-time market data, performs vector retrieval from the knowledge base, gathers macro and industry information, and outputs a structured data package; 2. Analyst: Conducts quantitative (valuation/technical indicators) and qualitative (competitive advantages/risks) analysis to generate investment arguments; 3. Critic: Performs logical consistency checks, bias identification, data integrity verification, and confidence scoring, then determines the process direction via conditional routing (if ≥ threshold → Decision, otherwise return to Analyst/Researcher); 4. Decision: Synthesizes the analysis to output buy/hold/sell recommendations and monitoring indicators.

The RAG implementation in PortfolioIQ is a key feature: Data sources include real-time market data (yfinance), historical analysis reports, and macroeconomic data; FAISS is used as the local vector storage (efficient similarity search, no external dependencies); the retrieval strategy uses a hybrid approach: semantic search, keyword filtering, and time decay to prioritize the latest reports.

Deployment: Docker containerization ensures environment consistency, and GCP Cloud Run is used to achieve automatic scaling and pay-as-you-go billing. Observability is implemented via LangSmith: execution tracing, latency analysis, error diagnosis, and cost tracking to facilitate debugging and monitoring.

The value of PortfolioIQ is reflected in multiple scenarios: 1. Individual investors: Automated research saves time, and structured arguments avoid emotional decisions; 2. Investment institutions: Assists junior analysts to improve efficiency, and standardized processes ensure output quality; 3. Quantitative strategies: Incorporates qualitative signals, generates event-driven trading signals, and monitors position risks.

Limitations include: High latency (not suitable for high-frequency scenarios), increased LLM call costs, hallucination risks (requiring manual verification), possible delays or missing data from free data sources, and the need to pay attention to regulatory compliance. Key takeaways: Role division optimizes tasks, conditional routing achieves a quality closed loop, RAG ensures data authenticity, structured output facilitates integration, and observability is essential for production. This architecture can be extended to fields such as legal analysis and medical diagnosis.