Laurium: Using Large Language Models to Extract Structured Data from Unstructured Text

Laurium is an open-source Python toolkit developed by the UK Ministry of Justice, specifically designed to extract structured data from free text and generate synthetic data using large language models. It supports multiple LLM backends such as local Ollama and cloud-based AWS Bedrock, can be adapted to different use cases via prompt engineering, and helps organizations unlock hidden value in text data.

In a data-driven decision-making environment, large amounts of unstructured text (such as customer feedback, support tickets, survey responses, etc.) contain valuable insights but are difficult to analyze quantitatively; traditional manual annotation is costly and cannot handle large-scale datasets. Laurium was developed by the UK Ministry of Justice's Analytical Services Team, originating from the BOLD Families project (aimed at estimating the number of children in the UK whose parents are in prison), to address the pain points of unstructured text mining.

Laurium can convert unstructured text into structured data, such as extracting sentiment tendency, urgency level, responsible department, and whether action is needed from customer feedback. Its application scenarios include customer feedback analysis, support ticket processing, research studies, public opinion monitoring, compliance review, etc., providing a foundation for quantitative analysis in data-driven decision-making.

Laurium adopts a modular architecture with core components including an LLM interface layer, prompt engineering module, output parser, and batch processing engine; it provides a dual-mode feature set (default core LLM functions + optional advanced machine learning functions), with a layered design to adapt to different user needs—allowing most users to get started quickly, while advanced users can enable advanced features for deep customization.

Laurium supports multiple LLM backends: local Ollama (no API cost, privacy protection, offline availability), cloud-based AWS Bedrock (powerful model capabilities), and is compatible with the LangChain ecosystem (e.g., ChatLlamaCpp), providing users with flexible choices.

Demonstrates a sentiment analysis pipeline: create an LLM instance, define the output schema, build prompts, generate a Pydantic model, extract and process data, and output structured results; supports simultaneous extraction of multiple fields (such as sentiment, urgency level, department, etc.), allowing multi-dimensional information to be obtained in a single LLM call, improving processing efficiency.

Installation methods: supports installation from PyPI/GitHub via uv or pip (with standard and advanced installation options); as an official UK Ministry of Justice project, it has trusted features such as production environment validation, security compliance, long-term maintenance, and openness/transparency, making it suitable for use in sensitive environments.

Laurium is a positive attempt by government agencies to open-source AI tools, encapsulating LLM capabilities into an easy-to-use, production-deployable solution, providing a reliable option for organizations that need to unlock the value of unstructured text.