What is Classification module?

The Classification Module in Supplier Data Manager helps you automatically assign products to categories and families based on your predefined structure. The AI Classification version uses AI to analyze data and classify products accordingly, reducing manual effort.

How does it work?

After mapping your supplier's data fields to your internal structure (cf. Mapping module), the Classification Module uses AI to evaluate each product and assign it to the most appropriate category or family. Most products will be classified automatically, but we recommend reviewing the AI predictions to ensure accurate classification.

Up to configuration, you decide whether you need to automate the classification or not: 

• Choosing automated mode means that the system will give you only the product rows to review that the AI wasn't able to classify automatically. The classified rows are visible from the “checked by AI” blue tab in SDM
• Choosing “to check”, means that 100% of the product rows need manual verification, even if AI has found a match for category or family. The classified rows are visible from the “to check” red tab in the SDM.

How to manually classify your products? 

You can manually classify products by adding one or more categories/families.

1. Select the row(s) that you want to edit.
2. Review the suggested category. Ifincorrect, use the search function to find a more accurate one.
3. Confirm the correct category.
4. Once all rows to check are validated, click "Finalize" to proceed.

How does AI classify products?

AI Classification uses pre-trained LLM (Large Language Models) to classify products into categories or families based on their attributes. It does not require additional training or user feedback loops to improve performance.
The model performs accurately out-of-the-box, eliminating the need for training datasets or ongoing model updates.

Using LLM models for classification offer

• Reduced Time-to-Value: Save up to 75% of the effort previously required for training-related tasks.
• Wider Compatibility: Supports multi-language classification (e.g., Japanese, German, Arabic).
• Improved Accuracy: Tested across diverse datasets to ensure robust performance in real-world scenarios.

[🤖 BETA] Tailor the IA to your context with custom prompts

The LLM doesn't know the specificity of your company context, your categories specificities, you can now add all the specific information in the custom prompt edition:

• Go to your project > Configuration > Project, and click the link to access the Configuration App.
• Select the relevant project and the classification step you want to customize.
• You'll find a text box where you can enter instructions for the AI to refine its classification process.

How to write an efficient prompt?

Custom prompts can help refine classifications by improving accuracy rather than enforcing strict rules. By optimizing your inputs and ensuring relevant information is present, you can guide the AI toward more precise and reliable results. 

To do efficient prompt we advise you to:

• Ensure that the sources provided to the AI contain sufficient information, such as names and descriptions, which often include key details
   
• Provide examples to improve the model’s ability to follow additional instructions. The more examples you give, the higher the chance the model will respect them.
   
• Provide labels rather than codes in your prompts to ensure accurate classification. Unlike text, codes need to be explicitly introduced in the prompt for the model to interpret them correctly.
   
• Provide the full path of your categories. For example, instead of only including “Wine fridge” in your prompt, provide the complete path: “Kitchen > Large appliances > Refrigerators > Wine fridge”
   
• Don't make LLMs predict something that contradicts their general knowledge.
  For example, if you make an LLM predict the category "refrigerator" when the product description clearly states that the product is an "oven", you'll have a hard time getting the model to predict "refrigerator" because it's inconsistent with how it was trained.
   

Common Questions (FAQ)

How does AI Classification agent understand our data?
LLMs are trained on extensive sets of text, allowing them to effectively interpret and classify product attributes from the beginning. SDM benefits from the best AI models on the market - GPT 4o mini
 

What if the AI makes a mistake?
LLMs models are designed for high accuracy but are not infallible. Misclassifications can be corrected during the step.
 

Can the AI learn from corrections?
No, AI Classification agent models don’t rely on feedback loops. As part of future developments, we want to enable adjustments to prompts.
 

How long does setup take?
AI Classification agent is the default model setup at SDM and can be used right away.
 

Is AI Classification agent available for all customers?
Yes, except for those requiring highly customized taxonomies or scenarios where the technology is unsuitable.

Limitations

• Maximum number of products (rows) per job: 100,000
• Maximum number of attributes: 50