Agentic Graph RAG: Self-Correcting Retrieval over Scientific Literature

TL;DR — A LangGraph agent over 2,000 arXiv CS papers that routes between vector, graph (with OWL ontology expansion), and community retrieval, rewrites failed queries, and recovers coverage from 27.5% → 81.2% through mode-aware self-correction.

[Live Demo] · [Architecture]

---

Overview

Standard RAG systems pick one retrieval mode and have no recovery mechanism when it fails. This work asks: can an agentic loop with mode-aware query rewriting recover queries that any single retrieval mode would refuse?

We build a knowledge graph from 2,000 arXiv CS papers (CS.AI + CS.CL, 2026) and wire three retrieval backends into a LangGraph state machine: vector search, graph traversal (with OWL ontology expansion), and community detection. When a context quality grader rejects the retrieved context, the agent rewrites the query in the vocabulary of the next retrieval mode and re-routes — up to three correction loops. A four-version ablation isolates the contribution of each component.

Key finding: adding a correction loop without query rewriting (v3) gives no coverage improvement over naive retrieval (27.5% vs 37.5%). Adding mode-aware rewriting (v4) recovers coverage to 81.2%. The gain is entirely attributable to rewriting, not to the loop structure or web fallback.

---

Results

80 queries · 30 factual / 30 relational / 20 thematic · GPT-4o-mini as RAGAS judge

Coverage across ablation versions

Version	Overall	Factual	Relational	Thematic
v1 — Naive vector	37.5%	93.3%	6.7%	0.0%
v2 — Static routing	28.7%	6.7%	36.7%	50.0%
v3 — Loop, no rewrite	27.5%	10.0%	16.7%	70.0%
v4 — Full system	81.2%	96.7%	53.3%	100%

v4 RAGAS by query type

Query type	Coverage	Faithfulness	Ans. Relevancy	Ctx. Precision
Factual	96.7%	0.966	0.890	0.907
Relational	53.3%	0.438	0.738	0.363
Thematic	100.0%	0.812	0.685	0.660
Overall	81.2%	0.789	0.789	0.697

Loop efficiency (v4)

Query type	Avg loops	1st-mode success	Router accuracy
Factual	0.28	82.8%	86.7%
Relational	0.63	62.5%	100.0%
Thematic	0.30	75.0%	55.0%
Overall	0.37	75.4%	83.8%

---

System

Retrieval mode	Backend	Best for
Vector	Qdrant hybrid (BGE-M3 dense + SPLADE sparse, RRF + cross-encoder rerank)	Factual, definitional
Graph + Ontology	Neo4j Cypher + fuzzy entity linking + OWL ontology expansion (130k triples, 47k inferred)	Relational, authorship, method categories
Community	BGE-M3 similarity over Leiden cluster embeddings + Groq summaries	Thematic, trend

Knowledge graph — 2,000 Paper · 9,250 Author · 3,003 Institution · 286 Method · 13 Community · ~50,000 edges

Ontology — 5 method subclasses (FineTuning · Attention · Alignment · Reasoning · Retrieval) · 82,573 explicit triples · 47,721 inferred triples · 29.9% paper coverage

Stack — Neo4j AuraDB · Qdrant · LangGraph · Groq LLaMA 3.3 70B · BGE-M3 · spaCy · rdflib · owlrl · FastAPI · Gradio

---

Ontology Layer

The OWL ontology is a classification layer built on top of the Neo4j knowledge graph. The knowledge graph stores explicit facts — which paper uses which method, who authored what. The ontology adds meaning — what category each method belongs to — and uses a reasoner to infer new relationships that were never stored explicitly.

Schema (ontology/arxiv_cs.ttl)

Five top-level classes: Paper, Author, Institution, Method, Community. Methods are further classified into a five-class subclass hierarchy:

Subclass	Examples
FineTuningMethod	LoRA, QLoRA, PEFT, Instruction Tuning
AttentionMethod	Transformer, FlashAttention, MoE, SSM
AlignmentMethod	RLHF, DPO, PPO, RLAIF
ReasoningMethod	Chain-of-Thought, RAG, LangGraph, GNN
RetrievalMethod	BM25, DPR, ColBERT, FAISS, Qdrant

The schema is written in OWL/RDFS Turtle format. relatedWork is declared as an owl:SymmetricProperty — if paper A is related to paper B, the reasoner infers the reverse automatically.

Build process (src/ingestion/build_ontology.py)

1. Export all nodes and edges from Neo4j as RDF triples
2. Classify 96 high-frequency methods into the 5 subclasses using Groq LLaMA 3.3 70B
3. Assert subclass membership (ex:method_LoRA rdf:type ex:FineTuningMethod)
4. Run owlrl.DeductiveClosure(RDFS_Semantics) to infer new triples — any two papers sharing a method subclass become relatedWork

Result: 130,294 total triples (82,573 explicit + 47,721 inferred) covering 598 / 2,000 papers (29.9%).

How it integrates with retrieval

Ontology retriever — queries the in-memory RDFLib graph via LLM-generated SPARQL. Handles questions about class membership and inferred relationships: "What category does LoRA belong to?", "How are DPO and PPO related structurally?"

Graph retriever (Option A expansion) — when a query mentions a category keyword (fine-tuning, alignment, reasoning...), the graph retriever first queries the ontology for all members of that subclass, then runs Neo4j Cypher for each — turning category-level questions into specific method lookups without hardcoding method names.

Known limitation

Method coverage is bounded by the 96 predefined METHOD_PATTERNS. Of 286 unique method strings in Neo4j, 62 matched and received subclass assertions. The remaining 224 exist in the graph but have no ontology classification. See ontology/SCOPE.md for details.

---

Quickstart

git clone https://github.com/VinaySampath14/agentic-graph-rag.git
cd agentic-graph-rag
pip install -e ".[dev]"
cp .env.example .env          # add Neo4j, Qdrant, Groq, Tavily keys
python scripts/verify_connections.py
python app.py                 # Gradio demo at localhost:7860

Run tests:

pytest tests/unit/ -v         # 44 unit tests, no credentials needed

See ARCHITECTURE.md for a full walkthrough of the codebase and CONTRIBUTING.md for setup details.

---

Citation

If you use this work, please cite:

@misc{vudumula2026agenticgraphrag,
  title   = {Agentic Graph RAG: Self-Correcting Retrieval over Scientific Literature
             via Mode-Aware Query Rewriting and Agentic Loop Control},
  author  = {Vudumula, Vinay Sampath Kumar},
  year    = {2026},
  url     = {https://huggingface.co/spaces/VinaySampath/agentic-graph-rag}
}

---

Related Work

Closest prior work: arXiv:2508.05660.
This work extends it with an agentic self-correction loop, mode-aware query rewriting, temporal graph edges, an OWL ontology layer with automated method classification and owlrl reasoning, and per-query-type RAGAS evaluation across a controlled four-version ablation.