Why Free Text Breaks Pipelines

Parsing

Part of: Structured Outputs

You ask a model for a price and it answers, "Sure! The total comes to about \$42.50 (plus tax)." A human reads that in a blink. Your code, trying to pull out a number, chokes. That one friendly sentence is why so many AI features work in the demo and fall apart in production. What it is A pipeline is any chain where a model's output feeds the next step automatically: a database insert, an API call, a UI that renders fields. Pipelines need machine-readable data, predictable shapes a program can index into. Free text is the opposite: fluent, varied, and unpredictable. The model that's so good at sounding human is, by default, terrible at being a reliable data source. The core mismatch: an LLM optimizes for plausible prose , but your code needs rigid structure . Those two goals fight every single request. How it works Walk through what breaks when you parse prose: 1. You ask for data. "Give me the customer's name and order total." 2. The model answers in a sentence. "The customer is Dana Lee and they owe forty-two fifty." 3. Your code tries to extract. It hunts for a name and a number, often with brittle string slicing or a regular expression. 4. The format drifts. Next request says "

Challenge: Count the parsing landmines