Definition
Zero-shot extraction is the ability to extract structured data from web pages or documents without providing any training examples specific to the target site or data format. The AI model relies entirely on its general understanding of language, web content, and the extraction instructions to identify and pull the requested information.
Zero-Shot vs Few-Shot vs Fine-Tuned
Zero-Shot
No examples provided. The model works from instructions alone: "Extract the product name, price, and description from this page." The model has never seen this specific site but understands the concepts well enough to extract the data.
Few-Shot
A handful of examples are provided (typically 2-5) showing input-output pairs. The model learns the pattern from these examples and applies it to new pages.
Fine-Tuned
A model is trained on hundreds or thousands of examples from specific sites or domains. Highly accurate but expensive to create and maintain, and limited to the trained domains.
Why Zero-Shot Matters
Immediate Deployment
Zero-shot extraction works on any page immediately. There is no data collection, labeling, or training phase. You point it at a URL, describe what you want, and get results.
Universal Applicability
Because it requires no site-specific training, zero-shot extraction scales to any number of target sites without proportional increases in setup effort.
Resilience to Change
Without site-specific training data to become outdated, zero-shot extraction is naturally resilient to site redesigns and content changes.
When Zero-Shot Falls Short
Highly specialized domains with non-standard terminology, complex nested data relationships, or unusual page layouts may benefit from few-shot examples to guide the model. Output quality can vary more with zero-shot compared to approaches trained on specific formats.
Zero-Shot Extraction in ScrapeGraphAI
ScrapeGraphAI's extraction engine operates in zero-shot mode by default. You provide a URL and a schema or prompt — no training data, no site-specific configuration. The AI generalizes from its broad understanding to extract the data you need, making it practical to start collecting data from new sources within seconds.