With AWS generative AI, Natura's Busca Inteligente serves one million consultants

Natura's Senior Cloud Engineer, Eli Melo, explains that the previous search system only allowed searches for exact terms, which could often discouraged purchases. “The person would type in a term and if there wasn’t an exact match among the products, there would be no result. The user was left without an answer and wouldn't buy,” she explains, citing the example of the Kaiak perfume, which the system couldn't find if the consultant typed in “kayaks.”

In addition, the search engine limited the number of characters per product name, which also hampered the search, since many products were registered with abbreviations. To meet the consultants' expectations, Natura decided that it was time to modernize the system using more sophisticated and efficient technologies that would make the search smarter.

“We started thinking about using algorithms and AI to carry out the search. During the study period, we contacted AWS, who provided us with an immersive experience in AI-based services,” she recalls. According to Melo, one of the products that appeared as a solution was precisely Amazon OpenSearch Service, which already includes built-in algorithms that allow for searches using context and phonetics, in addition to text.

The executive explains that in order for the search to be concluded as it should be, it was necessary to apply AI algorithms on the product database. Thus, the solution was found in existing algorithms, such as phonetic search, vector search and tag generation using Large Language Models (LLM), allowing for a smart search.

Hence the decision to develop an internal AWS services-based search solution, such as Amazon OpenSearch, which offers flexibility in storing product registration data and allows the application of a variety of Machine Learning (ML) algorithms, allowing phonetic, vector and other searches; and Amazon Bedrock, which allows payment on demand, eliminating the concern with creating and maintaining the infrastructure. “In addition to accessing a variety of available LLMs, Bedrock also allows us to take advantage of native integrations with other AWS services, which make the solution simpler to operate,” explains Melo, recalling that the LLM models also allowed the addition of terms (in the form of tags), relating the registered products with contextually pertaining terms.

The solution also used AWS Lambda, which with its serverless approach allows code execution without the need to manage servers, with automatic scalability and charging based solely on execution time. Melo points out that this combination creates a more flexible infrastructure, designed as follows:

“We were working on these ideas as a study, on the basis of trial and error. A cool point is that AWS allows us to easily upload things, test, validate, etc. A simple test doesn't take months. This speed helped us show value quickly,” he says. The executive explains that, in this way, a first non-production environment was developed, showing that the issues related to the previous system were resolved. So, the search system was improved until it was placed in a production environment, initially with 5% of the sales team, focusing on carrying out tests and collecting metrics. From the idealization to this first version, the entire development was carried out in one month.

Natura's Senior Cloud Platform Engineer & DevOps for Latin America, Marcelo Fabricanti, recalls that this first version was fully operational when delivered, allowing for improvements such as the inclusion of new metrics. “Today we map what people are searching for and are moving to a third phase: whether the search result went to the cart, and determining the page on which the search result was found. We're working on relevance. It's a long-term process,” he says.

Fabricanti points out that since the implementation of Busca Inteligente, there has been an increase not only in the results obtained, but in their accuracy. In Brazil, the solution is currently available to 100% of Natura's more than one million consultants, who perform approximately 400 searches per minute every day.

Compared to the previous system Natura has obtained results such as 205,000 searches with products added by the new system; 54,000 searches with results that would not be generated by the previous system; and 99% of searches with related products, added to the result. “We can conclude that the new search engine is not restricted to including products. It also generates results that, in many cases would not have been obtained by the previous system. This results in an increased probability of the user adding products to the cart, which consequently increases the possibility of conversion,” he says.

Fabricanti also points out that, when analyzing the performance of Busca Inteligente, his team found that, on average, queries are completed in under 50 milliseconds, which indicates functional improvements and performance gain compared to the previous system. “The idea for the future is to use the new mechanism on all of our channels: apps, end consumers, etc. Today we are in one of the applications, but there are other channels that can still receive this remodeled search. The idea is to expand to all channels,” he concludes.