TechWatch AI - Intelligent Monitoring and RAG Platform

TechWatch AI is a full-stack monitoring platform designed to automate technology watch. It leverages Retrieval-Augmented Generation (RAG) to provide a chat interface capable of answering queries based specifically on recently processed technical articles.
The project is built using a microservices architecture, combining a robust Spring Boot backend, a specialized FastAPI worker for AI tasks, and a modern Next.js frontend.

---

Architecture Overview

The system consists of three specialized services:

1. Java Orchestrator (Spring Boot)

• Role: Core business logic and data persistence.
• Features:
  • Automated RSS feed polling via Spring Scheduling.
  • Asynchronous processing of incoming articles.
  • Relational data management with PostgreSQL.
  • Centralized RESTful API for the frontend.

2. Python AI Worker (FastAPI)

• Role: Specialized AI processing and web scraping.
• Features:
  • Advanced Scraping: Content extraction using Trafilatura with BeautifulSoup fallback and boilerplate sanitization.
  • Vectorization: Embedding generation using the nomic-embed-text model via Ollama.
  • Semantic Search: High-performance vector retrieval using pgvector.
  • RAG Pipeline: Contextual answer generation using Large Language Models (Gemma / Qwen family).

3. Next.js Frontend

• Role: Interactive user dashboard.
• Features:
  • Real-time monitoring feed displaying extracted tech news.
  • Intelligent Chat Interface with Markdown rendering.
  • Responsive design with Tailwind CSS and system-aware dark mode.

---

Interface Preview

Real-Time News Monitoring Feed

RAG Assistant in Action

---

AI and RAG Implementation

TechWatch AI implements a Retrieval-Augmented Generation workflow to ensure response accuracy and minimize hallucinations:

1. Ingestion: Articles are scraped via Trafilatura to remove non-essential elements (ads, navigation, newsletters).
2. Embedding: Text is converted into 768-dimension vectors and stored in PostgreSQL using the pgvector extension.
3. Retrieval: User queries are vectorized to perform a cosine similarity search, retrieving the most relevant context snippets from the database.
4. Generation: The local LLM processes the query alongside the retrieved context to produce a grounded, professional response with Markdown formatting.

---

Technical Stack

Backend Orchestrator

• Language: Java 21
• Framework: Spring Boot 3.2, Spring Data JPA
• Database: PostgreSQL 16 with pgvector
• Build Tool: Maven

AI Worker

• Language: Python 3.11+
• Framework: FastAPI, SQLAlchemy
• AI Integration: Ollama (nomic-embed-text, Gemma / Qwen)
• Scraping: Trafilatura, BeautifulSoup4, Requests

Frontend

• Framework: Next.js (App Router), TypeScript
• Styling: Tailwind CSS, Lucide React
• Markdown: react-markdown

Infrastructure

• Containerization: Docker and Docker Compose
• Networking: Tailscale private mesh

---

Key Highlights

• Cross-Language Architecture: Demonstrates a scalable bridge between a Java enterprise backend and a Python AI service.
• Smart Truncation & Formatting: Implements custom logic to handle context windows cleanly and renders structured Markdown responses.
• Unified Database: Direct integration of semantic search within a relational PostgreSQL database via pgvector, eliminating the overhead of dedicated vector stores.
• Production-Ready Code: Focus on error handling, logging, and asynchronous task management.