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

1. The Research Problem

Graph-augmented Retrieval-Augmented Generation (GraphRAG) has emerged as a popular approach to multi-hop question answering across enterprise corpuses. However, standard GraphRAG architectures rely on exhaustive LLM extraction passes to detect entities, summarize relationships, and generate dense community hierarchies.

In enterprise production environments, this creates severe engineering and economic constraints:

• Prohibitive Indexing Overhead: Running LLM extraction across millions of tokens incurs massive API costs and hours of pipeline processing.
• Query Latency Bloat: Multi-hop traversal over dense graphs introduces latency that exceeds acceptable interactive thresholds (>2-5 seconds).
• The Density Paradox: Increasing graph edge density improves retrieval only up to an inflection point. Beyond this threshold, additional semantic edges introduce topological noise, diluting precision and causing context pollution.

"The Density Paradox reveals that more relationships in an enterprise knowledge graph do not equal better retrieval. Above an optimal information entropy threshold, dense edges behave as topological noise."

2. The MaxEntRAG Method

MaxEntRAG addresses these challenges by replacing exhaustive LLM-based graph extraction with an information-theoretic anchor selection algorithm:

1. LLM-Free Deterministic Indexing: Indexes source documents deterministically in sub-second time without making a single external model call.
2. Maximum Entropy Anchor Selection: Computes the information entropy of lexical tokens and entities across corpus spans. Only high-information anchors that maximally differentiate evidence paths are selected.
3. Source-Anchored Transitive Links: Retains direct, immutable provenance pointers back to raw document spans, ensuring that every retrieved fact has an auditable evidence chain.

3. Key Experimental Findings

Evaluated systematically across standard multi-hop benchmarks—including HotpotQA, MuSiQue, and GraphRAG-Bench:

• 90%+ Reduction in Indexing Cost: By eliminating LLM extraction calls during indexing, document preparation cost is virtually eliminated.
• Sub-100ms Retrieval Latency: Direct sparse index traversal over entropy-gated anchors delivers order-of-magnitude faster retrieval compared to graph clustering algorithms.
• Comparable or Superior F1 Retrieval: Avoids the precision drop caused by dense graph hallucinated edges, preserving clean evidence boundaries for the synthesis model.

4. Citation & BibTeX

5. Related Works & Articles

This research connects with accompanying work on real-time context compression:

• MaxEntRAG-Flow: Incremental Evidence Structures for Context Compression in Long-Context LLMs (Zenodo)
• From MIS to Operational Intelligence in Enterprise AI Systems
• The Modern Data Backbone: Semantics-First Architecture