{"slug":"nexthink-fine-tunes-a-7b-llama-model-on-amazon-sagemaker-to-translate-natural-language-into-nql-queries-for-nexthink-assist","url":"https://findausecase.com/use-cases/nexthink-fine-tunes-a-7b-llama-model-on-amazon-sagemaker-to-translate-natural-language-into-nql-queries-for-nexthink-assist","title":"Nexthink fine-tunes a 7B Llama model on Amazon SageMaker to translate natural language into NQL queries for Nexthink Assist","description":"Swiss software company Nexthink moved its NL2NQL tool, behind the Nexthink Assist AI agent, from prompt engineering and RAG to a LoRA fine-tuned 7 billion parameter Llama model on Amazon SageMaker. It built a golden evaluation dataset and used Claude 3.5 Sonnet on Amazon Bedrock as an LLM-as-judge for labeling. The source reports a 30% improvement in query accuracy and 80% reduction in token usage.","company":"Nexthink","industry":"Technology & Software","country":"Switzerland","aiCapabilities":["Generative AI","Conversational AI"],"technology":["Amazon SageMaker","Amazon Bedrock","Claude"],"deployment":"Public Cloud","problemStatement":"Nexthink's NQL query language is powerful but creates a barrier for non-technical users, while demand for DEX insights was growing among other IT teams, HR and employees. Prompt engineering and RAG left performance inconsistent for complex queries, and prompting a general-purpose LLM brought high token costs and slow responses over the full data model.","solutionApproach":"Nexthink fine-tuned a 7 billion parameter Llama model with LoRA on Amazon SageMaker to power NL2NQL, the tool behind the Nexthink Assist agent that translates natural language into NQL queries. It built a golden dataset of questions and queries, first annotated by employees through SageMaker Ground Truth and then labelled by Anthropic Claude 3.5 Sonnet on Amazon Bedrock, with the LLM also acting as a judge and humans validating samples.","businessValue":"After fine-tuning, the source reports a 30% improvement in query accuracy, an 80% reduction in token usage and support across Nexthink's entire data model. The source also notes the trade-off of increased complexity: fine-tuning needs substantial compute, high-quality datasets and dedicated engineering effort. Figures are as reported in the AWS blog post.","evidence":{"band":"high"},"sourceUrl":"https://aws.amazon.com/blogs/alps/how-nexthink-builds-enterprise-ai-agents-using-fine-tuned-llms-with-amazon-sagemaker","dates":{"publishedAt":"2026-10-21T05:47:29.387Z","publishedAtSource":"pipeline","updatedAt":"2026-10-21T05:47:29.387Z"},"license":"Open for reading and citing with a link to https://findausecase.com/use-cases/nexthink-fine-tunes-a-7b-llama-model-on-amazon-sagemaker-to-translate-natural-language-into-nql-queries-for-nexthink-assist. Bulk republication requires permission."}