scieee AI-readable full text Open interactive document viewer

From Questions to Data: Incorporating SoftwareIncorporating LLMs into RDM Software

Schlemmer, Alexander

Publisher
Zenodo
DOI
10.5281/zenodo.17935099
Repository
zenodo.org
Language
en

Abstract

LinkAhead is an open source research data management platform whose core functionality rests on a powerful semantic search engine. The engine exposes a purpose-built query language called CQL that resembles English, enabling both interactive exploration through a graphical user interface and programmatic access in automated workflows. Despite its intuitive design, mastering the query syntax, particularly for complex, multi-hop traversals of LinkAhead’s graph-structured metadata (e.g., reference and link queries), remains a barrier for non-technical users and hampers broader adoption.To lower this barrier, we have developed a dedicated query interface powered by a custom-trained Large Language Model (LLM). The model is fine-tuned exclusively on a corpus of LinkAhead query pairs, consisting of a natural language description and a formal query. The system turns questions that users enter into LinkAhead into searches, handling both easy lookups and complex connections, without users needing to learn the query language.In this presentation, we will describe the prototype’s design and present preliminary validation results that assess the model’s quality. This event was part of the Data Days Lower Saxony 2025 - Virtual Theme Day. The event is organized by the Lower Saxony Research Data Management Initiative (FDM-NDS). FDM-NDS is a joint project under the umbrella of Hochschule.digital Niedersachsen and is funded by zukunft.niedersachsen, a funding program of the Lower Saxony Ministry of Science and Culture (MWK) and the Volkswagen Foundation.

Full text

From Questions to Data: From Questions to Data: Incorporating LLMs into RDM Software Incorporating LLMs into RDM Software Data Days Niedersachsen 2025 – Virtueller Thementag 2025-11-26 Alexander Schlemmer What is LinkAhead? What is LinkAhead? A 290 A 121 Exp 002 Lab data (ELNs, ...) Processing, Data Analysis Numerical Simulations Experimental Results Prüfzertifikat gemessene Werte: Ist: 358.12 Norm-Abweichung: 21.83 nach DIN ISO PRÜFSTELLE Publications Ideas, Protocols, Internal Documents, SOP 1 2 3 4 LinkAhead Thinking Data Management Ahead demo.indiscale.com demo.indiscale.com Agile, semantic data management Agile, semantic data management ●LinkAhead initially developed at MPI-DS (Göttingen) ●Released as open-source (AGPLv3) in 2018 (gitlab.com/linkahead) ●Commercial support by IndiScale GmbH LinkAhead Thinking Data Management Ahead Where LinkAhead makes an impact Where LinkAhead makes an impact Data Model Data Model The LinkAhead Crawler The LinkAhead Crawler ELNFile: type: ELNFile match: ^presentation-dataset\.eln$ subtree: AuthorDataset: type: ROCrateEntity match_entity_type: Person match_properties: name: (?P<name>.*)$ "@id": (?P<url>.*)$ records: Author: url: $url name: $name Connect to data analysis software Connect to data analysis software 0 2 4 6 ϕ 0 2 4 6 g(ϕ) 2.0 2.5 3.0 3.5 4.0 4.5 radius [arb. units] NL to SQT Translator - Container Diagram Lindi Data Generator [system] LINDI Microservice [system] LinkAhead [system] «container» Data!Generator [CreatesajsonfilewithNL-SQTpairs thatisusedfortraining] «container» POV!Cache!Generator [Generatesajsonfilewith pre-generatedPOVsthatcanbeused toproduceproperSQTs] «container» Training [TrainanLLMfortranslationofNLto SQTusingtheQLoRamethod] «container» JSON!file!with!cache «container» Training!Data!JSON!file «container» Lindi!Microservice «container» WebUI «container» LinkAhead!Server «container» LLM!model «person» IndiScale!Developer «person» End!User «system» Open!AI!GPT2!Model creates uses uses createscreates uses sends!NL!querytranslates!NL!to!SQTgenerates!data fine-tunes!the!modelgenerates!the!cache evaluates!the!model queries downloads Data Generator Data Generator ●Input: natural language query ●Output: expected structured query (SQT) ●Generate variations using LLM JSON (excerpt) with training data Variations of NL Query Variations of NL Query ●Find machines where temperature is measured along with frequency ●All ProductionMachines tracking both temp & freq ●List ProductionMachines that measure temperature and frequency ●records for production machines that have both temperature and frequency data ●(…) ●Records of ProductionMachines with temperature measurements available ●Request to ollama: ●“Generate {{ N }} diverse English paraphrases of the query below that preserve the EXACT same intent and constraints. (…)” ●Find all records of type ProductionMachine which have a temperature_measured and a frequency_measured First validation results First validation results ●Current validation accuracy is approximately 87% ●High control over types of supported CQL/SQT queries. Iteratively increasing complexity. ●Training on own hardware: Nvidia GeForce GTX 5090 ●Main limitation: Errors in training data Outlook: ●Invalid queries ●Custom data models as context ●Chat component Thank You! Thank You! More Information: ●Project page on www.indiscale.com ●IndiScale website: www.indiscale.com ●LinkAhead repositories: gitlab.com/linkahead/linkahead-server/ ●LinkAhead demo: demo.indiscale.com/