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 (EasyChair Submission #2, 2026-03-17). Notification: April 24, 2026.
Authors: Madhulatha Mandarapu, Sandeep Kunkunuru
Summary
Empirical comparison of three approaches for LLM access to knowledge graphs: (1) domain-specific MCP tools with Cypher templates, (2) generic text-to-Cypher via schema-aware prompting, and (3) 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.
Files
- LaTeX:
research/arxiv/paper6_aidm_2026.tex - PDF:
research/arxiv/paper6_aidm_2026.pdf - Anonymous version:
research/arxiv/paper6_aidm_2026_anonymous.pdf