Large Language Models Enable End-to-End Data Integration Automation: A Research Breakthrough from the University of Mannheim

The latest research from the Web Science team at the University of Mannheim has realized the automation of the entire data integration process using large language models (LLMs). This solution achieved performance comparable to human experts in three real-world cases, while reducing the time cost per use case from 19 person-hours to 2 hours, with a cost of only about $9. The research covers four key steps: schema matching, value normalization, entity matching, and data fusion, bringing a revolutionary breakthrough to the field of data integration. The related results have been accepted by the Beyond SQL Workshop 2026 (co-hosted with ICDE 2026).

Data integration is a core challenge in modern data engineering. Enterprises need to integrate multiple heterogeneous data sources (such as music, game, and business data). Traditional methods rely on manual pipeline configuration and data annotation, which are time-consuming and labor-intensive. The University of Mannheim team proposed an LLM-driven end-to-end automated pipeline to address this pain point and achieve efficient and low-cost data integration.

Core research question: Can LLMs achieve human expert-level performance in data integration tasks while reducing costs? The pipeline built by the team includes four key steps:

1. Schema Matching: A single-prompt LLM method, inputting source column names/samples and target JSON Schema, achieving an F1 score of 1.0;
2. Value Normalization: A hybrid strategy (rule-driven processing for standard formats + LLM processing for categorical attributes);
3. Entity Matching: FAISS candidate selection + LLM active learning annotation + traditional ML matcher, with an average F1 score of 0.937;
4. Data Fusion: LLM-generated validation set to select optimal rules + RAG-enhanced variant, with the RAG version achieving an accuracy of 0.773.

The team selected three real-domain datasets to verify generality:

• Music Dataset: Integrating MusicBrainz/Last.fm/Discogs (37k+ records, 8 attributes);
• Game Dataset: Integrating DBpedia/Metacritic/sales data (74k+ records, 12 attributes);
• Company Dataset: Integrating Forbes/DBpedia/FullContact (14k+ records, 10 attributes). In terms of performance, the average F1 score for entity matching was 0.937, exceeding the baselines of manual configuration (0.894) and manual annotation (0.916); schema matching achieved a perfect F1 score of 1.0.

The automated pipeline significantly reduces costs:

• Time Cost: Approximately 2 hours per use case, a 90% reduction compared to the manual baseline (19+ person-hours);
• Economic Cost: A total of about $27 for three use cases ($9 per use case), using the GPT-5.2 model (pricing as of February 2026). This has great commercial value for enterprises that frequently perform data integration.

Current limitations: The accuracy of the RAG version for data fusion (0.773) is slightly lower than manual configuration (0.800), and human experience is still needed for complex decision-making scenarios. Future directions include: exploring more advanced LLM models, optimizing prompt engineering, researching multimodal data integration, and extending to real-time data stream processing.

Industry implications: LLMs can take on end-to-end data integration, and the role of data engineers will shift from manual configuration to architecture design and exception handling; enterprises can use this to reduce costs and accelerate data-driven decision-making. The team has open-sourced the complete code, case data, and pipeline outputs (GitHub), based on the PyDI framework, including execution scripts, Jupyter Notebooks, prompt templates, etc., to facilitate research reproduction and industrial applications.