The exponential growth of digital news content has created an urgent demand for intelligent systems capable of synthesizing, comparing, and extracting actionable insights from diverse sources in real-time. This paper presents a Large Language Model (LLM)-powered research tool that enables users to input multiple online news article URLs and pose questions to extract concise, accurate, and source-aware answers. The system leverages Retrieval-Augmented Generation (RAG) principles by combining Hugging Face’s transformer-based embeddings, FAISS for efficient semantic vector search, and LangChain to orchestrate the pipeline. To preprocess dynamic web content, we employ Unstructured.io loaders, while response generation is handled by the Groq-accelerated LLaMA-3 model. The tool provides explainable, context-aware answers with traceable sources, supporting journalistic, academic, and research use cases. Experimental evaluation confirms the system’s effectiveness, achieving an average response latency of 0.32 seconds in optimized configurations while maintaining high relevance and factual alignment. Our framework offers a highly adaptable and scalable approach to real-time insight extraction over heterogeneous, web-based news content.
Paper
Full text
LLM Powered News Research Tool for Insight Extraction
Semantic Scholar · 2025
Abstract
The exponential growth of digital news content has created an urgent demand for intelligent systems capable of synthesizing, comparing, and extracting actionable insights from diverse sources in real-time. This paper presents a Large Language Model (LLM)-powered research tool that enables users to input multiple online news article URLs and pose questions to extract concise, accurate, and source-aware answers. The system leverages Retrieval-Augmented Generation (RAG) principles by combining Hugging Face’s transformer-based embeddings, FAISS for efficient semantic vector search, and LangChain to orchestrate the pipeline. To preprocess dynamic web content, we employ Unstructured.io loaders, while response generation is handled by the Groq-accelerated LLaMA-3 model. The tool provides explainable, context-aware answers with traceable sources, supporting journalistic, academic, and research use cases. Experimental evaluation confirms the system’s effectiveness, achieving an average response latency of 0.32 seconds in optimized configurations while maintaining high relevance and factual alignment. Our framework offers a highly adaptable and scalable approach to real-time insight extraction over heterogeneous, web-based news content.