A Hybrid RAG-Based Framework for Improved Information Retrieval from Long Documents via Summarization and Question Answering

Abstract

We present an integrated system, which is a mix of summarization as well as Question Answering (QA) to boost the performance regarding information retrieval (IR) for long and domain-intensive documents. This work creates a synergistic relationship between both tasks to enhance retrieval efficiency and quality of answers, unlike previous studies focusing separately on either summarization or QA. The framework proposed is comprised of three stages. First, it involves summarization that is used for condensing documents, maintaining their key semantic content. Second, it employs a retrieval-augmented generation (RAG) architecture, where the all-MiniLM-L6-v2 model is used for the retrieval stage, while T5 and BART-Large-CNN models are utilized for the generation stage. Lastly, answers generated are put into comparison with a baseline system which works with complete document processing without summarization. Results at the NIPS dataset reveal that integrating summarization with QA significantly improves each summary informativeness and answer accuracy. To investigate the impact of summarization inside the retrieval pipeline, two complementary eventualities are explored. First, Proactive Summarization as an Indexing Tool (Summarization-First) plays summarization earlier than the indexing and retrieval stages to reduce record size and improve retrieval performance while keeping semantic relevance. Second, Query-Guided Contextual Summarization (Retrieval-First Summarization) applies summarization after the retrieval level, wherein only the retrieved relevant passages are summarized in step with consumer question for generating more targeted and context-conscious solutions. These two scenarios enable a comparative analysis of the way the position of summarization within RAG pipeline impacts retrieval efficiency as well as solution era quality. 

Received: 19 June 2026 | Revised: 13 August 2026 | Accepted: 14 September 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in Kaggle at https://www.kaggle.com/datasets/benhamner/nips-papers?select=papers.csv. 

Author Contribution Statement

Ahmed Sami Jaddoa: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & Editing, Visualization, Project administration. Jaber Karimpour: Conceptualization, Methodology, Formal analysis, Investigation, Resources, Data curation, Writing – review & editing, Supervision, Project administration. Pedram Salehpour: Conceptualization, Supervision.