Good data catalogues are essential for efective data sharing and discovery to cope with the rapid expansion of datasets and scientific literature available on the Web. In this paper, we present LLMDapCAT, an LLM-based metadata and data catalogue system that exploits Large Language Models (LLMs) and Retrieval Augmented Generation (RAG) for eficient data profiling, sharing, and exploration. We demonstrate how the system serves both data providers and consumers: on the one hand, it allows providers to automatically generate standardized and semantically accurate metadata from scientific papers using an LLM and RAG-based pipeline, and to publish the metadata in the catalogue system; on the other hand, it enables consumers to browse available datasets and explore them in chat-like Q&A sessions using an external LLM service. The system can be applied to curate custom domain-specific scientific databases that facilitate search, understanding, and exploration of domain-specific Demo: https://github.com/SINTEF-SE/LLMDapCat_Demo ∗Corresponding author.
A key challenge in the current era of information overload is the limited efectiveness of traditional
keyword-based search over scientific literature and shared datasets. With the exponential growth of
available publications and data, platforms for data discovery must go beyond simple keyword matching.
There is an increasing need for data discovery on the Web that leverages semantic matching, which
requires high-quality, semantically rich metadata to accurately describe published datasets [
Manual annotation of metadata is labor-intensive, often inconsistent, and varies significantly across
domains. Automated approaches augmented with human-in-the-loop feedback mechanisms emerges as
a promising approach to address these limitations. Recent advancements in generative AI—particularly
the development of Large Language Models (LLMs), including both general-purpose models such as
ChatGPT, LLaMA, and Mistral, and domain-specific models such as BioGPT [
Exploiting these recent advances, this paper presents the LLMDapCat web application, including a Streamlit1-based interactive user interface and two LLM/RAG pipelines for metadata generation and dataset exploration.
CEUR
ceur-ws.org
Our main contributions are as follows: • We introduce a data catalogue system designed for intuitive data sharing, featuring an accessible user interface (UI) and automated metadata generation using LLMs. The generated metadata is aligned with domain ontologies to ensure consistency and improve findability. • We present a method to curate customized, domain-specific scientific databases that support research exploration and analysis.
LLMDapCat provides a ChatGPT-like interface for dataset profiling and discovery. While the system is demonstrated using biomedical datasets, it can be extended to other scientific domains.
2https://www.ncbi.nlm.nih.gov/home/develop/api 3https://www.ebi.ac.uk/biostudies 4https://www.ebi.ac.uk/biostudies/arrayexpress 5https://europepmc.org/RestfulWebService
LLMDapCat provides two main user interfaces: • Provider View: Allows users to submit textual documents and automatically generate metadata using the LLMDap pipeline. • Consumer View: Enables users to search, explore, and query datasets using a ChatGPT-style
Q&A interface.
Metadata Schema: One important input to the system is the metadata schema that includes metadata ifelds defined according to domain ontologies to ensure semantic alignment and interoperability. The LLMDap will extract metadata from the scientific papers based on the information contained in the schema, consisting of metadata name, description, and value ranges.
This page allows users (data providers) to submit documents (e.g., research papers) for automated LLM-based metadata extraction with the following steps: 1. Provide input paper: Users can upload a PDF/XML file or input a URL/PubMed ID. 2. Select schema: Users can use a default or custom JSON schema to guide metadata generation. 3. Process paper: Clicking “Process Input” button sends the paper and schema to the backend LLM pipeline for metadata generation. 4. Review metadata: The generated metadata is displayed for user validation or edits. 5. Save results: When confirmed, the metadata is saved to the database and linked with the input document.
This page ofers dataset browsing and management functionality. It allows users (data providers or consumers) to browse, search and select datasets processed and indexed in the catalogue. After a dataset is selected for Q&A (for data consumers), it will initiate the Consumer View page. In addition, this page ofers an optional database utility function to allow users (data providers or system administrators) to initiate index-update in addition to the automatic index-update associated with uploading of datasets and their metadata.
1. Browse datasets: A paginated view displays key metadata for available datasets. 2. Search datasets: A search bar allows keyword-based filtering. 3. Select for Q&A: Users select datasets via checkboxes and initiate Q&A by clicking on the “Ask
Questions About Selected Datasets” button shown on the right image of Figure 3. 4. Update index: Users can press the “Rescan Directories & Update Index” button (above the Browse Datasets header in the left image of Figure 3) and start a background job to rescan and reindex datasets.
The configuration page (Figure 5) allows selection of LLM models, tuning of parameters (e.g., temperature, max tokens), and prompt template customization.
Semantic Technologies: Metadata fields are aligned with established ontologies to ensure that LLM outputs are both accurate and interoperable.
We have introduced LLMDapCat, a web-based application and LLM-backed pipelines for enabling FAIR
data sharing and exploration. Our approach uses RAG and LLMs to improve the quality and
trustworthiness of generated metadata and Q&A interactions. Quantitative evaluation has been conducted on
the proposed system to validate the performance of the pipeline and showed systemic improvement in
the annotation task [
In future work, we plan to integrate domain ontologies more tightly into the profiling process to extend and refine metadata schemas. This will enhance semantic discovery, accuracy, and coverage of domain-specific metadata. In addition, qualitative user evaluation is planned to validate the usefulness of the system.
Potential Impact: The proposed system facilitates semantic data discovery and exploration for researchers. LLMDapCat can also be used to build customized scientific metadata catalogues in any domain by tailoring the metadata schema and integrating with domain-specific APIs and ontologies.
The work is funded through the projects UPCAST (HE 101093216), enRichMyData (HE 101070284), and DataPACT (HE 101189771).
The author(s) used GPT-4 for grammar and spelling checks. The author(s) have reviewed and edited the content and take full responsibility for the publication’s content.