[GsoC 2026 ] Agentic GraphRAG #36 on STaRK-MAG with OpenVINO — Working System + draft Proposal

[1Suryansh1]

Hi @bharagha @14pankaj,

I'm Suryansh Verma , B,Tech CSE Student from Delhi Technological University. I have a working, benchmarked GraphRAG system running on Intel OpenVINO, and I'd like to share it as my warm-up contribution for Project 36.

My Laptop is intel AI PC with Intel Core Ultra 7 155H with integrated Intel Arc Graphics and NPU — all inference runs on-device via OpenVINO

Results

I evaluated the current system on the STaRK-MAG human-generated split with FULL ALL 84 queries in a fully local, zero-shot setting on 1,872,968 NODES . The current no explicit rerank run achieves Hit@5 = 53.57 with average latency of about 10.2 seconds per query.

human_eval is considered tougher than synthetic split and synthetic has thousands of queries.

System	Hit@5
BM25	41.67
ada-002	41.67
GritLM-7b	44.04
Claude3 Reranker	45.24
GPT4 Reranker	46.43
This system	53.57

the exact numbers here https://arxiv.org/html/2404.13207v3
A few points matter here:

• I have considered Hit@5 bcoz this tool task is to retrieve top 5-10 which will be given to other agent.
• My current system already outperforms the published GPT4 Reranker baseline on the same split , both on score and latency as avg latency is 23.43 sec as mentioned in paper and my system latency is 10.2 sec per query
• I see this as a strong result for a system that is fully local, zero-shot, API-free, and not finetuned.
• Newer frontier systems such as HybGRAG and GraphRAFT report higher STaRK-MAG scores on synthetic split , but they come from later evaluation settings, and GraphRAFT is explicitly a finetuned method rather than the same local zero-shot regime. and their VRAM requirements are higher and some involves GNN also.
• My goal in Project 36 is to close part of that gap with SLM-based anchor extraction, reflection, and better seed quality, while preserving the local OpenVINO deployment path.
• I have done ablations and updating my repo to further strength my direction.

Current System

I built a graph retrieval tool for agents, implemented and benchmarked on STaRK-MAG. The system is designed as a reusable retrieval backend for agentic workflows, not just a benchmark script.

Offline setup

• I precompute 384-dimensional OpenVINO embeddings for STaRK-MAG nodes using a bi-encoder.
• I store node text and embeddings in Neo4j with both vector and full-text indexes for hybrid seed retrieval.
• I also build a local CSR-style adjacency index for fast in-process neighbor lookup during traversal.

Online retrieval flow

• A query first goes through hybrid seed retrieval in Neo4j using dense retrieval, BM25/full-text, and Reciprocal Rank Fusion.
• The selected seeds are then expanded with an MCTS-based traversal engine that performs multi-hop exploration, backpropagation, and pruning.
• Candidate paths are scored with a trajectory-aware OpenVINO cross-encoder PRM that uses a sliding-window serialization of recent hops, which helps keep longer paths compatible with the model context window.
• Neighbor expansion uses degree-penalized semantic selection, which reduces hub-node domination during traversal.
• Results from multiple seed-specific search trees are merged into one ranked candidate list with trajectory evidence.

Engineering choices

• The runtime uses two graph layers: Neo4j for indexed seed retrieval and a in-memory CSR/NumPy graph store for fast local traversal.
• The entire query loop runs fully locally, with zero network calls during retrieval and graph search.

OpenVINO implementation

• The reranker supports NPU → GPU → CPU automatic fallback with recompilation, applies INT8 quantization on NPU , further indepth details could be found in repo

Project 36 Direction

My proposal is to package this retrieval core as a GraphQueryTool inside a LangGraph agentic pipeline. The traversal engine remains a deterministic retrieval tool in the loop, while OpenVINO SLMs handle query understanding, reflection, and answer synthesis.

Stage	Model	Device	Role
Anchor SLM	Qwen3-1.7B INT4	NPU	Schema-free entity extraction into structured JSON anchors
GraphQueryTool	Existing system	GPU	Anchor-primed seed retrieval, graph traversal, ranked candidates, evidence paths
Reflective SLM	Qwen3-4B INT4	GPU	Retrieval sufficiency check and retry routing with better anchoring
Generative SLM	Qwen3-7B INT4	GPU	Final answer synthesis from top candidates with citations if possible

Text-to-Cypher fails silently on schemas — the LLM hallucinates relationship names and returns zero results with no recovery. My Anchor SLM extracts schema-free JSON entities instead. It never generates Cypher. Schema knowledge lives in the retrieval layer, not the LLM.

Project 37 compatibility

( with @naitik-2006 fine tuning and fulfilling its requirement could be done , I have connected with him via linkedin , we can collaborate together towards converging solution with your guidance )
The three SLM layers can be fine tuned properly.

The graph traversal engine is deterministic — no fine-tuning required, keeping the gradient signal clean and targeted.

The same structure is also naturally extensible to multimodal retrieval because the seed-retrieval layer can ingest new embedding types without changing the traversal core.

Conclusion

I'd like to propose this benchmarked implementation as my warm-up contribution for Project 36. It already demonstrates working Neo4j integration, OpenVINO inference, million-node-scale graph traversal, and retrieval evaluation on STaRK-MAG, which I believe maps directly to the goals of the project.

The choice of model for SLM could be further decided and extended to support multimodal nature also.
I'd be happy to open a PR integrating this direction into the edge-ai-libraries ChatQnA pipeline.

Your suggestions and directions would be very helpful before I draft my final proposal

the image shows the PRESENT SYSTEM which would act as deterministic tool

the image below shows the proposed architecture

For Further clarification plz have a look at my architecture images

GitHub: https://github.com/1Suryansh1/graphRAG
I have added the mentors as collaborator and to guide for next steps and corrections ...

Suryansh Verma
B.Tech CSE
Delhi Technological University