6 min read

Enhance search relevance with Artificial Intelligence

Continuously evolving product search experience

Thousands of hundreds of products are available and ready to be found in many commerce stores nowadays. Thus, an efficient and accurate search experience is crucial to give shoppers a joyful journey. Products must be retrieved effectively, being the order of relevance one of the main distinguishing factors of a good commerce store.

Learning to rank search relevance ranking

Generative AI as a leading model

The way shoppers search for products online has evolved dramatically. The days of classic keyword-based searches seem to be gone, and the path is clear for an innovative hybrid search based on artificial intelligence (AI), in particular on Generative AI.

Generative AI is a type of artificial intelligence technology used to create a wide variety of original content, such as texts, images, audio, videos, simulations, etc., understanding and mimicking human thought processes. As the generative AI models are inspired by the human brain, they rely on neural networks which consist of interconnected nodes that process information to learn from it.

Generative AI starts by analyzing vast amounts of training data to better understand language patterns and nuances. Then, it learns to recognize these patterns and relationships within the data (structure of words, phrases, and sentences, and their relation to each other in specific situations) to generate content by predicting what comes next in a given context. Lastly, it continually learns from mistakes and adjusts its output to produce more accurate and contextually relevant content.

Empathy Platform constantly adapts and enhances its capabilities and features to present customers, brands, and shoppers with the most relevant and personalized product results. Empathy Platform leverages generative AI technology to enhance the overall user experience while preserving shoppers' data privacy by using a text-based machine learning model built on top of Mistral AI’s open-source large language model (LLM). Mistral AI is a pre-trained dense general-purpose foundation model with a strong commitment to ethical data provenance.


Being transparent about the training data used to teach models is the biggest challenge for companies creating training models for Generative AI applications. However, this European-based initiative has delivered the highest open-source semantic model complaining with EU regulations in terms of AI.

Data provenance in foundation models

Data provenance is a real concern for foundation model users since everybody is contributing to them with data that is so difficult to track. A trustworthy data provenance focuses on better transparency, explainability, and safety in data management and machine learning models. As not all data sources are trustworthy, Empathy Platform implements a human-centered, privacy-first approach, where the shoppers’ private information is not used to feed the models.

So, Empathy Platform responsibly leverages domain-specific datasets based on consent integrity, anonymous, and session-based customers’ interactions to fine-tune the semantic foundation model.


Check out how Empathy Platform works to protect privacy and integrity of data.

Learning-to-rank models applied to search relevance with gPlay

The applications of generative AI are expanding every moment. One of them participates in quick and accurate ranking of products since product positioning is so important when creating enjoyable and trustworthy search and discovery experiences. Thus, Empathy Platform applies generative AI to developing a learning-to-rank model (LTR) because this model learns how to rank relevant product results for a query thanks to data sets obtained from supervised and automated generated feedback.

Learning to rank consists of a machine learning technique that trains models to rank products from a catalogue according to their relevance to a query. In such a way, they can be applied in the development of ranking models for search systems, such as your commerce store’s search and discovery experiences.

The most important factor in determining the search system's effectiveness for shoppers is the overall relevance of results retrieved for a query. The evaluation of the system ranking quality is based on Text Retrieval Conference (TREC) relevance judgments, where a multi-level approach is used to categorize the quality of the ranking results as very relevant, relevant, or not relevant.

Want to know more about how judgments are generated?

The generation of judgments means providing feedback, which can given implicitly or explicitly:

  • Implicit feedback is the most common way to generate judgments in an automatic fashion, where signals like clicks, add2cart, checkout, no results, low results, etc. are used to label relevant documents based on human interaction with a commerce store. These signals are collected from anonymized, content-integrity shoppers’ interactions and deeply analyzed so the Empathy Platform’s offline evaluation mechanisms can serve as a crucial training ground to better understand and predict relevance.
  • Explicit feedback, driven by human intervention, complements the implicit one to avoid biases like the clicks only in documents the shoppers can see or the click position bias, where the first positions of a ranking always have more clicks independently of their relevance.

To further enhance commerce search experiences, Empathy Platform is shaping gPlay AI, a conversational and generative back office playboard for merchandisers. This playboard, connected with the Mistral AI foundational model, includes three different spaces: Explainability, Analytics, and Feedback. In the Feedback space is where explicit feedback is given to accurately determine the relevance of the products listed in the search engine results page(SERP) for a specific query.

Merchandisers can decide how the explicit feedback is given, manually or automatically when using the analysis provided by generative AI. Merchandisers can also check how this feedback has been given to control the performance of the generative AI in terms of accuracy of the evaluations when classifying the products' relevance and to adjust the automation to the business needs.

How explicit evaluation is given?

There are two ways of giving explicit offline evaluation based on how judgments are generated:

  • The manual evaluation that merchants can make in the back office tool Empathy Platform is exploring consists of manual human intervention to train the underlying LTR algorithms. The merchant determines through manual selections or evaluations whether the product results ranked by the LRT algorithms are relevant for the given query in regards to shoppers' intent. Harnessing the power of offline evaluations is key to training modern machine learning algorithms and data models without biases.

  • Merchants may prefer to make an automatic evaluation, which consists of an AI-based selection of product results ranked for a given query that is evaluated automatically by generative AI algorithms. They determine which products are more relevant for the query based on the training received. Merchandisers must review to adjust or validate this feedback afterward.

With these evaluations, data sets are created to train and fine-tune the search relevance algorithms that adapt dynamically by leveraging LTR mechanisms. This ensures that search results align with shoppers' preferences, ultimately improving the overall relevance of the product ranking.

The quality of product ranking leveraged by the LTR algorithms and the answers of the automatic evaluations given by generative AI are being improved thanks to fine-tuned domain-specific data sets to offer your clients a high-quality and relevant search experience.

But it doesn’t end here, Empathy Platform is enhancing the conversation with clients through the search box with Retrieval-augmented generation (RAG) technology. This technology isn’t trained, as it learns from new information provided as external sources of knowledge.