Migipedia relies on AI: Efficient overview of 1.5 million reviews

One million product reviews - a reason to celebrate, but also a problem. While those responsible at Migros celebrated this milestone, many customers simply felt overwhelmed by the flood of reviews. Today, with 1.5 million reviews, the need for an overview is greater than ever.

«At the moment, there is an extremely high number of reviews. You can't keep up with reading. So you don't read any more, you leave Migipedia on the left ...»

Frustrierter Migipedia-User, 2023

Our solution: AI-based summaries that concisely present relevant opinions in just a few sentences.

The idea: user-centered innovation

In a workshop with the Migipedia community, we came up with the idea of summarizing ratings. With the Migros AI Foundation and Azure AI, we realized this efficiently and in compliance with data protection regulations. After testing several models, we opted for GPT-4o Mini - a cost-effective and precise choice.

A technological highlight with added value

The solution creates precise summaries in four languages, which are updated according to a defined set of rules as soon as a new rating is entered. The feedback is overwhelming: 85% of users find the function "helpful". Transparency remains a core principle. All summaries are labeled, and a reporting function enables quality assurance through AI-supported pre-analysis and manual follow-up.

From hackday to go-live: a practical path to an AI solution

The path from the initial idea to productive implementation began with a Proof of Concept (POC). In the few developer days we had, the motto was: just get started and verify the idea. We followed a pragmatic approach to create benefits quickly without gold-plating the solution.

🚨 Attention to data quality: Get good data for your POC! You can only really verify your idea with "real" data.

Despite little effort, we in the team were all delighted with the result.

We dealt with the following topics on the way to go-live:

• Data protection & governance: The protection of user data was a key concern from the outset. It was clear to us that no personal information would be used in the prompts. In addition, the Migipedia rules were adapted to ensure that no personal data is stored in the reviews themselves.
• Monitoring: A very important topic! In addition to regular spot checks, we have outsourced monitoring to the users themselves - entirely in the interests of the community 😜. The reporting function allows users to flag problematic content directly, which ensures the quality of the summaries in a simple and effective way.

In January 2024, we were finally able to go live and make the new feature available to users.

The solution is shown schematically in the picture:

1. When a new valuation is written, a set of rules checks whether a new summary needs to be generated
2. If so, we retrieve the ratings from the database and send them to the AI Foundation using a prompt
3. The newly generated summary is then saved in the database

Wähe, Flade or Tüle after all?

Was our solution perfect? Of course not. Our beloved "Wähe" (or Flade?!) - a term that our language model did not recognize and therefore stubbornly rendered as "Whe". This led us to adapt the prompt to be product-specific and thus teach the model the correct product name.

This example shows that the solution was not yet perfect - but its benefits were already enormous. The users were prepared to accept an 80% solution as helpful instead of waiting for a flawless version.

The iterative approach The Migipedia community proved its worth: with every piece of feedback from the community, we were able to improve the algorithms and optimize the system. As a result, we have developed and continue to develop a solution that is well received by the Migipedia community.

Progress in the retail trade

This solution may not revolutionize the entire retail world, but it is a prime example of how AI can create real benefits. Migros Online and Digitec/Galaxus are already showing interest in using this technology for themselves.

And we are staying on top of Migipedia: we are already analyzing how we can also use AI for categorization or detailed analyses to further improve the platform.