Enhancing Accessibility to Semantic Search in LinkAhead using Large Language Models

Abstract

LinkAhead is an open-source research data management platform with a powerful semantic search engine. Queries are expressed in CQL (CaosDB Query Language), a domain-specific language that supports complex traversals of LinkAhead’s graph-structured data model. Despite its expressive power, CQL syntax remains a barrier for non-technical users. We present LINDI, a translation pipeline that converts natural language queries into valid CQL without requiring users to learn the query language. The pipeline decomposes translation into a sequence of focused Large Language Model (LLM) calls, producing a Query Intermediate Representation (QIR) that is validated against the active data schema before query execution. The QIR is human-readable and directly editable, giving users a transparent checkpoint to inspect and correct before a query runs. We evaluate LINDI on 200 manually compiled queries across four datasets (three in English, one in German) from independent research domains. On the three English-language datasets, Llama 3.1 (8B) achieves consistent accuracy between 90.1% and 97.7%. GPT-OSS (20B) ranges from 91.5% to 93.2%. Overall accuracy across all four datasets is 88.0% and 85.0% for the two models respectively, and the pipeline’s modular design allows the LLM component to be replaced independently. A single-step baseline that skips this decomposition reaches only 70.5%/72.5% overall accuracy on the same queries, indicating that decomposing translation into schema-aware, LLM-assisted steps with human review is a practical route to making LinkAhead’s semantic search accessible without requiring users to learn CQL.