Adaptive Chunking and Model-Aware Retrieval for Multi-Hop RAG

This repository contains the implementation for the research paper "Adaptive Chunking and Model-Aware Retrieval Strategies for Multi-Hop Retrieval-Augmented Generation".

It implements a Model-Aware RAG pipeline that dynamically adjusts chunk sizes to optimize retrieval for both token-constrained and large-context embedding models.

📄 Abstract

Existing RAG systems often rely on fixed chunking strategies that fail to account for model-specific constraints. This project introduces an adaptive pipeline that:

• Token-Constrained Models: Uses 1500-character chunks for models like BERT/BGE to prevent truncation.
• Large-Context Models: Uses 5000-character chunks for models like Qwen/Snowflake to maximize semantic coherence.
• Structure Preservation: Converts JSON corpora to PDF to retain layout and hierarchy before segmentation.

🚀 Key Features

• Adaptive Chunking: Logic to switch between chunk sizes based on the selected embedding model.
• Hybrid Retrieval: Combines dense vector retrieval (OpenSearch) with sparse retrieval (BM25).
• Multi-Hop Reasoning: Optimized for questions requiring evidence from multiple documents (MultiHop-RAG benchmark).
• Hallucination Reduction: Larger context windows reduce evidence fragmentation, leading to lower hallucination rates.

📂 Repository Structure

• app.py: Main application entry point (Streamlit interface).
• JsonToPdf.py: Preprocessing script to convert JSON datasets into structure-preserving PDFs.
• split.py: Core logic for the adaptive chunking strategy (sentence-aware segmentation).
• docker-compose-opensearch.yml: Docker setup for the OpenSearch vector database.
• MultiHop-Rag/: Dataset and core processing modules.

🛠️ Getting Started

Prerequisites

• Python 3.8+
• Docker (for OpenSearch)

Installation

1. Clone the repository:

   git clone [https://github.com/DS-ProjectRepo/rag.git](https://github.com/DS-ProjectRepo/rag.git)
   cd rag

2. Start the Vector Database (OpenSearch):

   docker-compose -f docker-compose-opensearch.yml up -d

3. Run the Application:

   python app.py

📊 Performance Highlights

Based on the paper's evaluation:

• Retrieval: Significant improvements in Hits@10 and MAP@10 using adaptive chunking.
• Generation: Reduced hallucination rates for inference-heavy queries when using 5000-character chunks with large-context models.

---

For more details, please refer to the associated research paper.