RAGNews — Agentic LLM-Powered News Platform

Overview

RAGNews is a production-grade agentic news platform that ingests, indexes, and intelligently answers questions about news using a Retrieval-Augmented Generation (RAG) pipeline. The system uses a domain-fine-tuned LLM for summarization and a BERT-based classifier for sentiment analysis.

Pipeline Architecture

1. Agentic Routing

Built on FastAPI + LangChain, the system routes user queries through database-aware agents that autonomously decide between vector retrieval, keyword search, or direct answer paths.

2. Two-Stage Retrieval

• Stage 1 (Coarse): Vector search over Milvus for fast approximate nearest-neighbor retrieval.
• Stage 2 (Fine): Cross-encoder BGE-Reranker re-ranks top candidates for precision. This decoupling minimizes latency while maintaining quality.

3. LLM Fine-Tuning

• Fine-tuned Qwen2.5-1.5B using LoRA + Unsloth for domain-specific news summarization.
• Reduced training memory footprint by 60% vs. full fine-tuning.
• Integrated BERT-based sentiment classifier for financial news polarity tagging.

4. Production Serving

• Fine-tuned models served via vLLM for high-throughput inference.
• PostgreSQL for short-term and long-term conversation persistence.
• Streamlit UI for interactive sessions.
• Full Docker containerization for deployment.

Tech Stack

System Architecture

The two-stage retrieval pipeline:

1. Milvus Vector Search: Fast approximate nearest-neighbor retrieval from indexed news corpus
2. BGE-Reranker: Cross-encoder refinement for precision ranking

Demo

The platform provides:

• Interactive question-answering interface
• AI-powered news insights and sentiment analysis
• Real-time article retrieval and summarization

Key Results

• 60% reduction in GPU memory for fine-tuning via LoRA + Unsloth
• Asynchronous concurrent data ingestion pipeline handling multiple news feeds
• Low-latency retrieval via coarse–fine two-stage architecture