Overview

With hundreds of thousands of new products added annually to its digital marketplace, the multinational retailer needed a way to efficiently manage the generation and review process for new product descriptions. Leveraging AI21 Labs’ Jamba model, AI21 Solution Architects collaborated with the retailer’s team to build a custom solution that supports the retailer’s key metrics of accuracy and latency, ultimately improving accuracy by up to 95% on classification and text generation tasks, while reducing latency for workflow processing that’s 7X faster than competitors.

Opportunity

The multinational retailer is a European retail giant specializing in electronics and entertainment devices, beauty products, and household appliances. The company operates a network of physical stores and e-commerce platforms, offering a diverse range of products and services to customers across various regions.

Through the company’s marketplace, third-party sellers can list and sell their products alongside the retailer’s own offerings. Each year, third-party sellers launch hundreds of thousands of new products; overall, the company manages millions of products, each with its own description and attributes. Amidst such a high volume, manually sampling and reviewing these descriptions to ensure adherence to corporate guidelines and quality standards was proving to be unsustainable and hard to scale.

Solution

To overcome this operational efficiency challenge using AI, while maintaining extremely high quality standards, AI21 Solution Architects built a solution that carefully orchestrates a number of steps to extract, validate, and generate product information, resulting in a final outcome that is trustworthy and aligned with the company’s internal brand guidelines.

Defining Product Description Attributes

First, AI21’s in-house team of professional linguists collaborated with the multinational retailer to define a set of quality criteria for product descriptions. Criteria spanned both the accuracy of the descriptions (i.e. do the words accurately describe the picture of the product?), as well as their structure and language, optimizing to ensure maximum clarity and readability. The AI21 linguists helped translate and map the retailer’s desired attributes for marketplace product descriptions into measurable LLM outcomes. Not only does this step serve as scaffolding for the solution’s architecture, mapping quality criteria to LLM outcomes ensures that appropriate mechanisms are in place for reviewing and auditing the AI solution’s decision-making.

Assessing Submitted Product Descriptions

Using this mapping, the AI21 Solutions Architect team developed corresponding architecture and custom classifiers to evaluate and classify product descriptions as meeting—or not meeting—the defined quality criteria. This step operationalizes the retailer’s quality criteria for evaluating product descriptions submitted by third-party sellers, a core component for enabling the automation of the submission review process. Using these custom classifiers, the solution flags products that do not meet the quality requirements, as well as identifying the exact content within their description that does not align with the retailer’s selected quality criteria, further streamlining the review process.

Automatically Generating Product Description Text

For submissions flagged to the retailer’s quality assurance team, the AI21 solution automatically generates a new description, now in alignment with the company’s quality criteria, to ensure that only approved product descriptions are added to the marketplace.

Efficiently Handling Large Request Volumes

Given the high volume of incoming requests on a daily basis, AI21 Solution Architects sought a way to ensure the turnaround time for the submission review and approval process would be kept to a minimum, an especially important criteria for maintaining a frictionless experience for the sellers. To solve this challenge, they collaborated with AI21’s engineers to implement a Batch API, which enables users to submit multiple requests at once for asynchronous processing, rather than handling requests individually with immediate responses. 

Yet instead of a standard Batch API—which usually batches requests just for offline tasks—AI21’s engineers built a Batch API that could batch requests for online tasks. Given the retailer’s product description volume, this custom-built Batch API proved integral to deploying an AI solution that could sustainably keep pace with demand.

Outcome

At the conclusion of the work process, the multinational retailer received an end-to-end API for handling both classification and generation tasks, saving them both time and resources on solution development and maintenance.

Additionally, the solution met the retailer’s key metrics of accuracy and latency, with Jamba reaching up to 95% accuracy across classification and text generation tasks, and the implementation of the custom Batch API reducing latency by 7X, as compared with traditional rate limits.

Between the success of the AI solution and the close strategic partnership developed between AI21 Labs and the multinational retailer, the customer opted to expand the collaboration, applying AI21’s solution to a wider volume of their product description data.

Architecture diagram

AI21 Tech & Services Used

  • Jamba Foundation Models: AI21’s Jamba Foundation Models leverage their hybrid Transformer-Mamba architecture and 256K context window to efficiently and accurately process large volumes of data, such as regulatory guidelines and technical manuals. As open models, they can be securely deployed in a customer’s VPC or in air-gapped, on-premise environments.
  • AI21 Batch API: AI21’s custom-built Batch API enables the AI solution to submit multiple requests at once (“batching”) for asynchronous processing of online tasks, rather than handling requests individually or batching requests just for offline tasks.
  • Custom AI Systems: Dedicated teams of experts from AI21 work hand-in-hand with the client at every step of the way, from use case identification to solution architecture, and from data preparation to post-implementation iteration and troubleshooting.