Optimizing chemical risk documentation with LLMs

Sammanfattning

This research presents an innovative solution based on artificial intelligence, whichis used for verifying data in databases. The system aims to detect errors, includingduplicate rows, empty fields, or inaccurate chemical product descriptions storedin the database. The project was primarily designed to support industrial safetyassessments by automating data verification through comparisons with externalsources using the GPT-4 Turbo language model. The solution incorporates advanced prompt engineering techniques, includingretrieval-augmented generation (RAG), where validated feedback examples aredynamically retrieved and inserted into the model prompt. This approach enablesthe system to adapt to domain-specific knowledge without requiring fine-tuning.Simulated professional testing demonstrated measurable improvements in errordetection and risk prioritization, with accuracy rising from 53,3% to 83,3% afteriterative prompt refinement and feedback-driven retrieval. The project addresses challenges such as reducing LLM hallucination, ensuringfeedback reliability, and meeting safety-critical application requirements. It offersa scalable foundation for semi-automated chemical documentation validation inindustrial environments.In conclusion, this demonstrates that artificial intelligence can be integrated in orderto improve data quality and thus minimize the human burden in environmentsthat require both accuracy and speed.