📈 Agentic RAG for Time-Series Analysis

An end-to-end multi-agent Retrieval-Augmented Generation (RAG) system built to provide a conversational interface for complex time-series data. This architecture seamlessly merges traditional data engineering (Airflow, dbt, Postgres), machine learning (ARIMA, XGBoost, Isolation Forests), and modern AI orchestration (LangGraph, OpenAI) to dynamically answer complex questions about past metrics, future forecasts, and contextual anomaly explanations.

---

🏗️ Architecture Overview

The system is designed in a highly modular, full-stack architecture:

graph TD
    classDef db fill:#3366ff,stroke:#fff,stroke-width:2px,color:#fff;
    classDef agent fill:#ff9900,stroke:#fff,stroke-width:2px,color:#fff;
    classDef pipeline fill:#00cc66,stroke:#fff,stroke-width:2px,color:#fff;
    classDef ui fill:#e11d48,stroke:#fff,stroke-width:2px,color:#fff;

    subgraph Data & ETL Pipeline
        A[Airflow Ingestion with yfinance]:::pipeline -->|Real Market Data| B[(PostgreSQL Relational)]:::db
        B --> C[dbt Staging & Marts]:::pipeline
        C --> D[(PostgreSQL Marts)]:::db
    end
    
    subgraph Machine Learning Hub
        D --> E[Isolation Forest Anomaly Det.]:::pipeline
        E -->|Triggers| F[Gemini Context Generator]:::pipeline
        F -->|Embeddings| G[(pgvector Index)]:::db
    end
    
    subgraph LangGraph Multi-Agent System
        U[FastAPI Backend]:::ui --> H{Master Router}:::agent
        H -->|Numerical Queries| I[SQL Agent]:::agent
        I -->|Queries| D
        H -->|Predictions| J[Time-Series Agent]:::agent
        J -->|Fetches Data| D
        H -->|Anomaly Context| K[Vector RAG Agent]:::agent
        K -->|Similarity Search| G
        
        I --> L[Synthesis Agent]:::agent
        J --> L
        K --> L
        L --> M[Natural Language Response]
    end
    
    subgraph Web UI
        W[React / Vite Frontend]:::ui <--> U
    end

Loading

---

⚙️ Technical Details

Phase 1: Data Storage (pgvector)

We use a unified PostgreSQL 16 instance as both a traditional analytical warehouse and a Vector Database.

Phase 2: ETL Pipeline (Airflow & dbt)

• Apache Airflow: Orchestrates daily ingestion of real stock market data (Apple, Google, S&P 500) via yfinance.
• dbt (Data Build Tool): Transforms raw data into staging and mart models, generating critical ML features (rolling averages, 1h/24h lag variables, standard deviations).

Phase 3: Time-Series Modeling Hub

A dedicated Python module (time_series_hub.py) housing standard algorithms:

• ARIMA: For baseline univariate forecasting.
• XGBoost: For multivariate forecasting utilizing dbt-generated lag features.
• Isolation Forests: For unsupervised anomaly detection. Anomalies trigger Gemini to write a short contextual summary, which is embedded via GoogleGenerativeAIEmbeddings and saved to pgvector.

Phase 4 & 5: LangGraph Agents & Full-Stack UI

A stateful, multi-agent workflow orchestrated via LangGraph, exposed through a Web UI:

1. FastAPI Backend: Provides the API layer for the agentic workflow.
2. React (Vite) Frontend: A premium, dark-mode web application featuring real-time chat and dynamic Recharts graphs overlaying historical data with AI forecasts.
3. Agents: Master Router, SQL Agent, Time-Series Agent, Vector RAG Agent, and Synthesis Agent work together utilizing Gemini 1.5 Flash to analyze data and synthesize conversational responses.

---

🚀 Getting Started

Prerequisites

• Docker & Docker Compose
• Python 3.10+
• OpenAI API Key

Installation

1. Start the Database:

   # Spins up PostgreSQL with pgvector and initializes the schemas
   docker compose up -d

2. Install Dependencies & Seed Data:

   pip install -r requirements.txt 
   
   # Run the standalone script to download real stock data into your database
   python seed_data.py

3. Setup Environment Variables: Ensure you have a .env file with your GOOGLE_API_KEY.

4. Run the Backend & Frontend: Start the FastAPI backend:

   uvicorn backend.main:app --reload

   In a new terminal, start the React frontend:

   cd frontend
   npm install
   npm run dev

---

Generated as part of the Agentic RAG for Time-Series Analysis architecture build.