Text-to-Cypher

Text-to-Cypher is the process of automatically translating natural language questions into executable Cypher queries using large language models. For example, the question "What projects is Alice working on?" would be converted to MATCH (p:Person {name: 'Alice'})-[:WORKS_ON]->(proj:Project) RETURN proj.name, proj.status.

Implementation requires providing the LLM with the graph schema (node labels, relationship types, property names), few-shot examples of natural language to Cypher translations, and rules about query safety (preventing destructive operations like DELETE or REMOVE). The LLM uses this context to generate syntactically valid Cypher that matches the graph's actual structure.

Production text-to-Cypher systems include query validation (checking syntax and safety before execution), self-correction loops (if a generated query fails, the error message is fed back to the LLM for a retry, typically with 2-3 attempts), and result formatting (converting raw query results into human-readable answers). This technique is a key enabler of hybrid RAG+KG systems, allowing the query router to access structured graph knowledge without requiring users to learn Cypher.