How to Build an Extraction Model with Promptrepo?

This is a common question from Google Sheets users who copy details out of text by hand. For example, you may copy ingredients and nutritional values out of food descriptions. It is slow and each page writes them differently. A simple formula cannot find them in free text and a general AI model does not return them in the same shape each time. With Promptrepo, you can build an extraction model from your examples and it will return the same fields every time. Even better, your training data stays in Google Sheets, so the people who know your data can build the model without coding.

Let's start. Create your AI model and follow the steps below.

Extraction models identify and extract structured information from raw text. In this tutorial, we'll build a model that extracts ingredients and nutritional values from food descriptions.

Step 1: Preparing Training Data

For extraction, our dataset consists of:
  1. Nutrition per serving: The nutritional values per serving
  2. Web page content: The contents of the web page
  3. Web page title: Title of the page
  4. Ingredients: The extracted list of ingredients
  5. Nutritional category: Categories such as beverage, red meat, etc
  6. Serving size (grams or mL): The food amount per serving
You can start with one example, as below and add the rest to its Google Sheet later.

Step 2: Creating the model in Promptrepo

  • Once you click on Create your AI model, you will land on the "Create your own AI model" page
  • Under Input, click on Text. You can also click on Document or Email.
Input with Document, Email and Text to choose from, with Text marked
  • Enter one example in the box: a web page title and its content, such as "Classic beef burger. A quarter pound patty with onions and cheese in a soft bun."
  • Click Done. It takes about 20 seconds and the Output on the right will be filled in by AI.
Enter Text with the example typed in the box, with the Done button marked
  • "Editing input" will be displayed. Click + Add input field if you need another input.
  • Click Done
Editing input on the left and the Output filled in by AI on the right, with the Done button marked
  • "Editing output" will be displayed. Check that the output fields are the ones you need, such as Ingredients and Nutrition per serving. Click + Add output field for any that is missing.
  • Click Done
Input on the left and Editing output on the right, with the Done button marked
  • Click Publish
Input and Output side by side, with the Publish button marked at the top
  • Enter the AI model name and click Publish Now
The AI model name box with Food extraction typed in, with the Publish Now button marked
  • "Model is ready!" will be displayed. Click Proceed.
The Model is ready! message, with the Proceed button marked

Step 3: Training & Testing the Model

  • Click PREVIEW
The top of the model page with EDIT, CODE and PREVIEW, with PREVIEW marked
  • Enter a food name and its description and click PROCEED. Enter it first, because PROCEED with an empty box shows an error.
  • The model should extract structured data (ingredients & nutrition) from the text
  • Click ADD TO TRAINING DATA on a result you like, to teach the model with one more example
The Input and the Output the model gave, with the Add to training data button marked
To add many examples at once, on the model page, click on the edit icon on the Training data tab. Its Google Sheet will open and you can add rows there.
The Training data tab of Food extraction-V1 with the example in a row, with the edit icon marked
Extraction models enable AI to understand and process textual data. Now, let's move on to generative models to create new data based on learned patterns.

Summary

If you copy details out of text by hand, it is slow and the pages all write them differently. With Promptrepo, you enter an example, review the output fields and publish an extraction model. Create your AI model and complete the steps above. If you have any questions, write to support.

Last updated: