Wove case study, continued

In Chapter 6, Gathering Data – Content is King, the Wove case study on data cleansing for rate sheets used by freight forwarders was kicked off. They had to scrub the data before ingesting it to output a clean, unified view of all the rates from many carriers. Now, it is time to explore their prompt engineering (we covered the basics in Chapter 7, Prompt Engineering) and fine-tuning efforts for this solution.

Prompt engineering

They want the LLM to think like a customer who does this step manually. They created the prompt for the spreadsheets during the ingestion process (this early version was shared with us to maintain the proprietary nature of their latest efforts):

You are an expert at table understanding. You will be given a snippet of text and a row number for the header row. Your task is to determine where the data for the table starts and what range of rows make up the header for the table. If the header has ambiguous columns (such...