Cdiscount is slowly but surely committing to generative AI

Artificial intelligence within Cdiscount is not new. Isabelle Serot, Data Director of the e-retailer, puts the number of “large algorithms” already in production at 50. To improve them, Cdiscount is focusing in particular on generative AI, she confided on February 6 during the Hub Retail & e-commerce conference.

The French actor did not wait for ChatGPT to take an interest in IAGen’s models. Experiments were carried out from summer 2022 on OpenAI LLMs. “They were great, but they didn’t outperform our internal AI models. We were already very advanced in AI, which I fondly call classical AI.”

Effective combination of AI and IAGen

A change of models is therefore ruled out during this period due to lack of gains in terms of performance, and for a price which is also more expensive then. Cdiscount’s “playing field”, characterized in particular by 80 million products and 16 million monthly visitors, nevertheless makes it an ideal candidate for generative AI.

“It was almost easy to go there,” comments Isabelle Serot. However, Cdiscount has chosen “to be very careful” in the exploitation of these models. The online store therefore opted for an approach consisting of combining existing “very proven” systems and LLM.

The result: “an operating cost which is not very significant”, but this time with performance gains “which we could not achieve without the LLMs.” In terms of use cases, Cdiscount mainly focused on product sheets, “the crux of the matter” according to the Data Director.

“Bringing together the customer and the right product also comes down to semantics. And if the product sheet is not correctly written, does not have the right properties, the right visuals… Then it will be more complex to achieve it,” she explains.

14 million recategorized product sheets

With ChatGPT, Cdiscount has therefore expanded existing algorithms. In June 2023, the seller put into production on Azure (via a location in Europe) a first use case of generative AI on product recategorization.

On February 2, 2024, Cdiscount had submitted 35 million product sheets to the AI ​​- sheets for which the internal AIs estimated that possible improvements were possible. Of this total, 14 million were actually recategorized.

For what benefit? “When a product sheet is recategorized, it increases the conversion rate by 30%,” figures Isabelle Serot. During the same period, Cdiscount put into production, with iAdvize, a second use case: a pre-sales advice chatbot.

To control hallucinations, the robot only draws information from product sheets. And Cdiscount claims 70% customer satisfaction on the chatbot, “the equivalent of a conversation with a human advisor.”

2024, year of conversational

The company sees “a great future in conversation” thanks to generative AI. This area of ​​application also constitutes “the great ambition of 2024”. However, the Data management intends to go there “with great caution and in stages”.

And for Isabelle Serot, the improvements generated on the product sheets prepare the ground for the deployment of conversational. Other use cases on the files have been deployed with a view to improving internal efficiency.

Conversational and products are nevertheless closely linked, emphasizes the Data Director of Cdiscount. “We are improving the quality of semantics on both sides. This way, meeting love will be even easier.”

In order to encourage the emergence of other projects, the e-retailer has also opted for the acculturation of its 2000 employees to IAGen. A webinar, bringing together Data Scientists, was organized to present the essential technical themes, but also the opportunities and risks.

Github Copilot and Copilot 365 deployed

An IAGen Committee has also been set up responsible for evaluating and prioritizing use cases reported by the professions. “To stimulate ideas and desires”, a movement was initiated internally in the form of competitions and ideation workshops.

In terms of buying, Cdiscount has equipped 400 to 500 IT and Data employees with Github Copilot. As of January 1, Copilot for Office 365 was also deployed to 300 people “from very diverse professions, mainly users of PowerPoint, Excel, etc.”

Isabelle Serot hopes, at an unspecified date, to increase the number of Copilot 365 licenses to 1,000. In 2024, Cdiscount will continue to improve its existing algorithms and “roll out the conversational boulevard”, but with caution.

A multi-LLM and open source strategy

“The more we expand, the greater the risks of hallucinations,” she declares to explain this cautious approach of IAGen. These hallucinations represent “the great anxiety” of the company. “We see a lot of what is being deployed. We don’t want it.”

Cdiscount does not intend to lock itself into the OpenAI environment. Isabelle Serot announces an “assessment of the place” given to open source. Isabelle Serot imagines the future in the form of “a mix” of solutions, including Azure OpenAI, AzureML with open source and also with in-house models.

“On our internal infrastructures, we already have powerful GPUs. And we just ordered even more powerful ones. We will therefore have LLMs internally. It is important to play on this diversity,” she concludes.

Source link -97