LLM – Driven Data Extraction
Customer
Swiss Chemical Company
Converting semi-structured recipes and flavor data into a more accessible database format to improve selection procedures
Challenge
- Poor application of recipe data due to inconsistent formats (un- to semi-structured)
- Converge data points to one format and make available in one database, with thousands of files in scope
Solution
- Developed a prompt workflow to extract relevant data from internally deployed GPT instances and transfer it into a document database
- Developed a streamlit app and workflow for human validation and corrections to improve the quality of extraction
- Simplifying the validation workflow and postprocessing of the extracted data
Business Impact
- Integrated hundreds of existing recipes into a newly created database.
- The database will be a milestone in the flavor division
- Opens up discovery of other ways for data capturing and integration that can be used in the future.