Paper 6: MCP Tools vs Text-to-Cypher — aiDM @ SIGMOD 2026
Title: Domain-Specific MCP Tools vs. Generic Text-to-Cypher: How Graph Databases Become the Data Layer for AI Agents
Venue: aiDM Workshop at SIGMOD 2026, Bengaluru, India (May 31, 2026)
Status: Submitted
Authors: Madhulatha Mandarapu, Sandeep Kunkunuru
Summary
Empirical comparison of three approaches for LLM access to knowledge graphs:
- Domain-specific MCP tools with Cypher templates
- Generic text-to-Cypher via schema-aware prompting
- Raw LLM with no graph access
Evaluated on two benchmarks across biomedical and industrial domains.
Key Results
| Approach | AssetOpsBench (139) | BiomedQA (40) |
|---|---|---|
| MCP Tools | 99% | 98% |
| Text-to-Cypher (NLQ) | 83% | 85% |
| GPT-4 / 4o standalone | 65% | 75% |
The “inverted LLM” thesis: structured data + deterministic tools outperforms unstructured LLM reasoning on domain-specific factual questions.
Manuscript sources, reviewer feedback, and reproduction artifacts live in the private samyama-research repository under papers/paper6-aidm/.