Search engine for discovering works of Art, research articles, and books related to Art and Culture
ShareThis
Javascript must be enabled to continue!

Maximum Entropy Source-Anchored Sparse Retrieval with LLM-Free Indexing for Graph-Augmented RAG

View through CrossRef
Graph-augmented Retrieval-Augmented Generation (RAG) commonly relies on dense knowledge-graph construction, where entities and relations are extracted from source documents and materialized as a retrieval structure. While this can improve relational access, it also introduces high indexing cost, graph growth, and noisy traversal through over-connected semantic hubs. This paper presents MaxEntRAG, a sparse source-anchored retrieval framework that uses Maximum Entropy (MaxEnt) filtering for LLM-free indexing and LLM-free graph construction. Instead of using an LLM to construct an exhaustive synthetic knowledge graph, Max-EntRAG builds a compact transitive retrieval structure around maximum-entropy anchors. Source spans act as grounded hyperedges, sentence and paragraph boundaries provide local context, and probabilistic traversal decides how far retrieval should expand for a given query. The method is designed to preserve the minimum useful connectivity required for multi-hop evidence retrieval while avoiding low-information edges that increase graph density without improving answer relevance. The paper studies this trade-off through the Density Paradox: beyond a critical density, additional semantic connectivity can reduce retrieval precision by dispersing query signal across plausible but irrelevant paths. Experiments on GraphRAG-Bench (Medical and Novel subsets), HotpotQA, and MuSiQue show that sparse maximum-entropy anchoring can approach strong graph-based retrieval baselines while avoiding LLM-based graph construction and maintaining substantially lower indexing cost, graph density, and query latency. The results suggest that many RAG workloads do not require complete LLM-generated knowledge graphs; they require compact, source-grounded, high-information retrieval structures that can be indexed and connected without LLM extraction.
Elsevier BV
Title: Maximum Entropy Source-Anchored Sparse Retrieval with LLM-Free Indexing for Graph-Augmented RAG
Description:
Graph-augmented Retrieval-Augmented Generation (RAG) commonly relies on dense knowledge-graph construction, where entities and relations are extracted from source documents and materialized as a retrieval structure.
While this can improve relational access, it also introduces high indexing cost, graph growth, and noisy traversal through over-connected semantic hubs.
This paper presents MaxEntRAG, a sparse source-anchored retrieval framework that uses Maximum Entropy (MaxEnt) filtering for LLM-free indexing and LLM-free graph construction.
Instead of using an LLM to construct an exhaustive synthetic knowledge graph, Max-EntRAG builds a compact transitive retrieval structure around maximum-entropy anchors.
Source spans act as grounded hyperedges, sentence and paragraph boundaries provide local context, and probabilistic traversal decides how far retrieval should expand for a given query.
The method is designed to preserve the minimum useful connectivity required for multi-hop evidence retrieval while avoiding low-information edges that increase graph density without improving answer relevance.
The paper studies this trade-off through the Density Paradox: beyond a critical density, additional semantic connectivity can reduce retrieval precision by dispersing query signal across plausible but irrelevant paths.
Experiments on GraphRAG-Bench (Medical and Novel subsets), HotpotQA, and MuSiQue show that sparse maximum-entropy anchoring can approach strong graph-based retrieval baselines while avoiding LLM-based graph construction and maintaining substantially lower indexing cost, graph density, and query latency.
The results suggest that many RAG workloads do not require complete LLM-generated knowledge graphs; they require compact, source-grounded, high-information retrieval structures that can be indexed and connected without LLM extraction.

Related Results

Automating Information Retrieval from Biodiversity Literature Using Large Language Models: A Case Study
Automating Information Retrieval from Biodiversity Literature Using Large Language Models: A Case Study
Recently, Large Language Models (LLMs) have transformed information retrieval, becoming widely adopted across various domains due to their ability to process extensive textual data...
DARE-RAG: Difficulty-Aware Retrieval Expansion for Retrieval-Augmented Generation
DARE-RAG: Difficulty-Aware Retrieval Expansion for Retrieval-Augmented Generation
Retrieval-Augmented Generation (RAG) systems face a fundamental trade-off: query expansion can improve retrieval effectiveness for ambiguous or underspecified queries, yet indiscri...
A Review on Indexing Techniques and its application in Multilingual Information Retrieval System
A Review on Indexing Techniques and its application in Multilingual Information Retrieval System
To implement the indexing in multilingual dataset, the indexing process must know. This paper gives the brief about indexing and presents role of indexing, logical view of indexing...
Exploring Large Language Models Integration in the Histopathologic Diagnosis of Skin Diseases: A Comparative Study
Exploring Large Language Models Integration in the Histopathologic Diagnosis of Skin Diseases: A Comparative Study
Abstract Introduction The exact manner in which large language models (LLMs) will be integrated into pathology is not yet fully comprehended. This study examines the accuracy, bene...
JADE: jawbone lesion diagnosis and decision supporting system
JADE: jawbone lesion diagnosis and decision supporting system
Abstract Objectives To develop and evaluate JADE, a proof-of-concept retrieval-augmented generation (RAG) diagnostic assi...
A Systematic Literature Review of Retrieval-Augmented Generation Implementation for Enhancing Large Language Models in Education
A Systematic Literature Review of Retrieval-Augmented Generation Implementation for Enhancing Large Language Models in Education
The rapid advancement of Large Language Models (LLM) has led to the creation of increasingly adaptive intelligent learning systems. However, many educational implementations of LLM...
Self-updating Retrieval Mechanisms for Scalable and Personalized RAG Systems in Dynamic Environments
Self-updating Retrieval Mechanisms for Scalable and Personalized RAG Systems in Dynamic Environments
Retrieval-augmented generation (RAG) has become a practical mechanism for improving the factuality, traceability, and domain relevance of large language model (LLM) systems. Howeve...

Back to Top