Natural Language to Gremlin Queries with Amazon Bedrock

The AWS article details an innovative framework designed to translate natural language queries into Gremlin, the complex query language for graph databases, leveraging Amazon Bedrock models, particularly Amazon Nova Pro. This technology addresses the significant barrier specialized graph query languages pose, enabling non-technical users like business analysts and data scientists to seamlessly interact with and extract insights from intricate, interconnected data structures without requiring deep technical expertise.

The core methodology unfolds in three stages. Initially, the system meticulously extracts both structural graph knowledge—comprising vertex/edge labels, properties, and one-hop neighbors—and pertinent domain knowledge, which includes customer-defined constraints and LLM-generated semantic descriptions. Subsequently, this information is structured into a schema, similar to text-to-SQL processing, to enhance the model’s comprehension of graph topology. This stage involves question processing for entity recognition and context enrichment, coupled with a context generation component that integrates element properties, graph structure, and domain rules.

The final stage involves Amazon Nova Pro generating the Gremlin query. A key technical feature is its robust, iterative refinement mechanism: if a query fails execution, an error parsing system analyzes the issue, and the LLM regenerates and refines the query until it successfully aligns with database constraints. The framework’s efficacy was validated through extensive evaluation using an LLM-based judge (Anthropic’s Claude 3.5 Sonnet on Amazon Bedrock) against 120 queries, achieving an overall execution accuracy of 74.17%. Performance analysis revealed that the generated Gremlin queries boast execution latencies comparable to human-written queries, signifying no additional overhead. Furthermore, Amazon Nova Pro demonstrably superior query generation latency and cost-efficiency compared to benchmark models, making it a powerful and cost-effective solution for complex graph data interactions. This approach successfully navigates challenges like heterogeneous properties and complex graph structures by integrating graph and domain knowledge, employing Retrieval Augmented Generation (RAG) for query plan creation, and featuring an iterative error-handling system.

(Source: https://aws.amazon.com/blogs/machine-learning/generate-gremlin-queries-using-amazon-bedrock-models/)