Scholarlink

As the volume of scientific literature continues to grow increasingly fast, researchers face the challenge of staying up to date with relevant findings. This work presents a research paper summarizer that uses Retrieval-Augmented Generation (RAG) integrated with a local Large Language Model (LLM). The system combines semantic search with generative capabilities to produce concise, informative summaries of lengthy scientific papers. The presented work uses LangChain as the orchestration framework, ChromaDB for vector-based document retrieval, and Ollama to run LLM locally. This architecture ensures efficient data handling and full offline functionality. The RAG model first retrieves contextually relevant segments from a paper and then generates summaries conditioned on these segments, resulting in coherent and context-rich outputs. This approach improves comprehension, reduces cognitive load, and eliminates the dependency on cloud-based APIs, ensuring greater control over data privacy. The proposed solution achieved ROUGE-1, ROUGE-2, ROUGE-L and BERTScore of 43.09, 21.47, 40.08 and 85.25 respectively.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Retrieval-Augmented Generation Pipeline for Summarizing Research Papers Using Local LLMs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Francis Fernandes,
  • Anirudh Hanchinamani,
  • Tarun S. Bagewadi,
  • Ratan Dhane,
  • Sharada K. Shiragudikar

摘要

As the volume of scientific literature continues to grow increasingly fast, researchers face the challenge of staying up to date with relevant findings. This work presents a research paper summarizer that uses Retrieval-Augmented Generation (RAG) integrated with a local Large Language Model (LLM). The system combines semantic search with generative capabilities to produce concise, informative summaries of lengthy scientific papers. The presented work uses LangChain as the orchestration framework, ChromaDB for vector-based document retrieval, and Ollama to run LLM locally. This architecture ensures efficient data handling and full offline functionality. The RAG model first retrieves contextually relevant segments from a paper and then generates summaries conditioned on these segments, resulting in coherent and context-rich outputs. This approach improves comprehension, reduces cognitive load, and eliminates the dependency on cloud-based APIs, ensuring greater control over data privacy. The proposed solution achieved ROUGE-1, ROUGE-2, ROUGE-L and BERTScore of 43.09, 21.47, 40.08 and 85.25 respectively.