Products from Brochures

If you want to try this extractor, you can download it here:

​

Document Input Examples

Here are some examples of the types of documents that this extractor can process.

​

Example Output

This extractor will produce JSON data that looks like the following:

[
    {
        "name": "Bottle of red wine",
        "original_price": 14.99,
        "discounted_price": 9.99
    },
    {
        "name": "Bottle of white wine",
        "original_price": 12.99,
        "discounted_price": 6.99
    }
]

​

Extractor

Here’s a template for a product brochure extractor with some explanations for why some settings are chosen.

​

JSON Schema

Here’s the example JSON schema to extract product data from brochures. It extracts a list of products, each with a name, original price, and discounted price. The schema also contains validation rules for each field to ensure the extracted data is accurate and consistent.

​

Extraction Strategy

​

Strategy: parallel-auto-merge

For supermarket brochures with many independent products, the parallel-auto-merge strategy is the most effective. It provides fast extraction times and ensures all products are extracted, even from large brochures.

​

LLM Recommendation: google/gemini-2.0-flash-lite

For product extraction from brochures, google/gemini-2.0-flash-lite is a good and cost-effective choice. It provides a good balance between performance and cost, and is well-suited for extracting data from structured documents like brochures.

​

Context Settings

​

Chunk Size: 15k-50k

A chunk size of around 15.000-50.000 tokens is sufficient for product brochures, as each chunk typically contains multiple independent product listings. We don’t want to use a chunk size that is too large, as it may cause the extractor to miss some products.

​

Include Text: true

Including text is essential for extracting product names, descriptions, and prices.

​

Include Embedded Images: false

Embedded images are not necessary for extracting product data from brochures and can be excluded to reduce token usage and processing time.

If we wanted to, we could enable this to allow the LLM to assign product images for ERP systems or online shops.

​

Include Page Screenshots: true

Page screenshots are helpful if the input data is not OCR’d or if the layout of the brochure is important to determine the prices.

​

Mark Images with IDs: false

Since we are not associating images with specific product data in this example, marking images with IDs is not needed.

Next Steps