Post-processing turns a language model’s raw answer into usable results. For text extraction, the model returns a text string, so additional processing must recover attribute values and place them in a structured table. section 4.3.1
A reliable extraction workflow separates generation from cleanup:
- Define the attributes, source text, delimiters, and missing-value placeholder in the prompt. section 4.3.1
- Extract the table region between <BeginTable> and <EndTable> so preambles and explanations do not enter the data. section 4.3.3
- Split the region into rows by newline symbols and fields by pipe symbols, then trim fields and discard empty values. section 4.3.3
- When several answers are generated, evaluate or execute them and discard unsuitable results, such as SQL queries with syntax errors. section 3.5.3
Extracting applicant data
Given an email stating that Martin studies Computer Science at Stanford and has a 4.0 GPA, the model can return | Martin | 4.0 | Bachelor of Computer Science |. Post-processing converts this delimited text into a row such as ["Martin", "4.0", "Bachelor of Computer Science"].
