The AiPi Knowledge Base: Semantic Framework and Terminology for Interdisciplinary Research Programmes

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper examines how to address key challenges in developing a custom GPT (AiPi) intended to function as a “philosophical assistant” for knowledge integration in interdisciplinary research programmes. The AiPi (Artificial Intelligence Principal Investigator) is based on an LLM but is required to handle specific tasks involving complex philosophical concepts and to interoperate with external diagnostic tools beyond the model's native capabilities, such as the philosophical tool EPR (Epistemological Profile Radar). This makes the overall endeavour particularly challenging. In response to well-documented limitations of custom GPTs –including instability in instruction-following, limited use of previous contextual information, selective utilisation of the knowledge base, and “hallucinations”– the parametrisation of AiPi followed a cyclical process of design, training, evaluation, revision and re-evaluation, with iterative optimisation, prompt engineering, evals and supervised fine-tuning. Particular emphasis was placed on the design and content of a strictly structured knowledge base, organised into nine thematic files in a vector store, eight corresponding to the EPR variables and a ninth dedicated to the semantic framework and key terms of interdisciplinary integration, such as epistemological stances, cognitive and semantic alignment, the epistemic virtue index, and related terms. A key element in AiPi’s effective functioning is the use of a stable JSON string as the interface between the EPR diagnostic tool and AiPi, allowing standardised parsing of results, comparison with the “Epistemic Virtue Index” of each EPR variable, and the generation of concise, structured feedback. Beyond its analytical role, AiPi also functions effectively as a philosophy tutor and facilitator, supporting users’ education and the understanding of their individual worldviews within interdisciplinary collaborations. Its successful pilot deployment makes the future integration of AiPi into a web-based platform, in conjunction with EPR and an additional philosophical tool, Dynamic Ontology, possible, illustrating that overcoming the limitations of custom GPTs is primarily a methodological rather than a purely technical challenge.