Pranav Yadav's HG-RAG paper, published on arXiv April 16, 2026, proposes graph-traversal retrieval over hierarchical knowledge graphs. On worlds up to 800 nodes, it consistently beats flat dense retrieval on hierarchical, relational, and multi-hop queries according to the arXiv preprint.
Key facts
- Submitted to arXiv on April 16, 2026.
- Evaluated across 18, 200, and 800 node worlds.
- Four query types: local, hierarchical, neighborhood, multi-hop.
- Single author: Pranav Yadav, affiliated with MIT.
- No code or dataset released with the preprint.
Standard RAG systems retrieve from flat document stores using dense vector similarity, which fails when queries require hierarchical or relational reasoning across structured knowledge. HG-RAG solves this by treating the knowledge base as a directed graph with parent, child, and neighbor relationships.
The pipeline first resolves a named entity anchor from the user query, then expands context upward through parent nodes, laterally through relational neighbors, and downward through child nodes when needed. This yields a structured context window that preserves the graph's hierarchy, unlike the flat top-10 chunks returned by dense retrieval.
Evaluation Across Three World Scales
Yadav evaluated HG-RAG against a dense retrieval baseline across three world scales—18, 200, and 800 nodes—using four query types: local fact, hierarchical, neighborhood, and multi-hop. HG-RAG consistently outperformed the baseline on hierarchical, relational, and multi-hop reasoning tasks while reducing hallucination and maintaining locality coherence.
The paper does not disclose absolute accuracy numbers, only relative improvement claims. No code or dataset is linked in the preprint, making independent replication impossible today.
Why Graph-Aware RAG Matters
Enterprise knowledge graphs—used in supply chains, biomedical ontologies, and organizational hierarchies—contain exactly the kind of structured relationships that flat RAG mishandles. A query like "Which cities trade with Voss?" requires traversing trade relations between cities, not just retrieving documents containing "Voss." HG-RAG's graph-traversal approach mirrors how a human analyst would navigate a database: follow edges, not cosine similarity.
The single-author paper from MIT (Yadav's affiliation, per the arXiv metadata) suggests this is early-stage academic work. The 800-node maximum world scale is small; real-world knowledge graphs often exceed millions of nodes. Scaling HG-RAG to production sizes would require hierarchical indexing and caching strategies not addressed in the paper.
What to watch
Watch for a follow-up preprint or code release from Yadav scaling HG-RAG to million-node graphs, or a benchmark comparison against GraphRAG (Microsoft's 2024 approach). Enterprise adoption will depend on whether the traversal overhead can be amortized at production scale.
Source: arxiv.org
