NFO — LLM Ontology Generator

• Source on GitHub

Problem

Building a domain ontology starts with competency questions: structured queries that define what the ontology must be able to answer. Writing these questions is expert-intensive and time-consuming. The goal was to use LLMs to accelerate early-stage ontology development without sacrificing domain specificity.

Role

Solo backend author. Responsible for LLM integration, prompt engineering, API design, authentication, and deployment.

Approach

• Users specify a domain and scope; the system generates a candidate set of competency questions using OpenAI GPT via LangChain
• An iterative prompt design process (few-shot examples, chain-of-thought) reduced hallucinations and improved question quality
• A validation layer checks structural consistency before returning results
• Google OAuth2 authentication gates API access to authorised researchers

Stack

• Language: Python 3.8+
• Framework: Flask
• LLM orchestration: LangChain + OpenAI GPT
• Auth: Google OAuth2
• Dev environment: Nix (reproducible)

Key decisions

• LangChain abstracts the LLM provider — swapping models requires minimal code changes
• Interactive user inputs (domain + scope) kept prompts grounded and reduced generic outputs
• Separated question generation from validation so each step can be improved independently