Skip to main content
T-Brain is Topsort’s commerce AI platform. It helps retailers power search, recommendations, product ranking, and predictions using models tailored to their products and shoppers. You can use T-Brain without using Topsort’s advertising platform.

​
What is a Large Commerce Model?

A Large Commerce Model, or LCM, is a generative AI model built to understand commerce. It learns from product information, shopper behavior, and context to understand what shoppers want and which products are relevant. For example, an LCM can help find products that match a shopper’s request, rank recommendations, or predict how likely a shopper is to purchase a product.

​
How is an LCM different from a Large Language Model?

A Large Language Model, or LLM, primarily learns patterns in language. An LCM learns from commerce signals, including product catalogs, shopper interactions, and purchases. This helps T-Brain account for both what a product is and how shoppers interact with it.

​
What can I use T-Brain for?

T-Brain supports four main capabilities:
  • Retrieval: Find relevant products for search results and recommendations.
  • Prediction: Estimate outcomes such as product clicks and purchases.
  • Ranking: Order organic products, sponsored products, or both according to the outcome you want to optimize.
  • Understanding: Generate embeddings, numerical representations of products and shoppers that capture similarities and relationships. You can also use these in your own models.
These capabilities can support search, website personalization, communications, and commerce agents.

​
Do I need to be an existing Topsort customer?

No. You can upload data from your existing systems, train models, evaluate them, deploy them, and call them through an API without using Topsort’s ad server. Existing Topsort customers can also use datasets already synced from Topsort.

​
What data do I need?

Provide your product catalog and at least one month of event history, including all purchases and paid product clicks and impressions. T-Brain organizes data into three dataset types: categories, products, and events. Events must include a consistent shopper identifier so interactions and purchases can be connected. If you are a Topsort customer, you will already have three datasets synced:
  • Topsort-categories
  • Topsort-products
  • Topsort-events
These datasets update in real time. You can use them when creating training tasks and upload additional datasets when needed. See Data Requirements for the file format and column requirements.

​
How do I upload data?

See Data Requirements for instructions on creating datasets, uploading Parquet files, and adding updated data.

​
How do I train a model?

A task defines what you want a model to do and which datasets it should learn from.
  1. Review the available models in the Model Garden tab. Choose the model and capability appropriate for your use case, such as product ranking or conversion prediction.
  2. Prepare your datasets in the Data tab. Use synced Topsort datasets or upload your own files.
  3. Create a task in the Tasks tab. Choose the capability, select the datasets it requires, and name the task. For example, a product-ranking task uses product and event datasets.
  4. Start training. T-Brain trains the model using the selected datasets.
  5. Follow progress in the Training tab. Review queued, running, and completed training jobs.
  6. Review the results in the Evals tab. Inspect model-quality metrics and compare performance before choosing to deploy.
  7. Deploy the model. Your application can then call it through an API.
  8. Test the deployed model in the Playground tab. Try requests and inspect the returned predictions and response times.

​
What does fine-tuning mean?

Fine-tuning adapts a base model using your retailer’s data. It specializes the model for your catalog, shoppers, and use case. T-Brain’s base model provides a starting point. Fine-tuning helps it learn the patterns specific to your store.

​
How can I evaluate a model?

Use the Evals tab to assess model quality and compare trained models with a reference model or baseline. You can also provide baseline data to compare your existing approach with T-Brain. Evaluations can include:
  • Model quality: How well a model performs its prediction or ranking task. For example, AUC ROC measures how well a prediction model distinguishes positive outcomes from negative ones.
  • Business impact: Comparisons such as sales value per session, conversion rate, advertising revenue, and return rate.
  • Performance: Response time, including P95 latency, the time within which 95% of requests complete.
The Evals tab includes evaluations, benchmarks, test sets, and experiments. Use the measures relevant to your task and the data available for the comparison.

​
What safeguards and controls are available?

T-Brain provides several points of review and control:
  • Data validation: Uploaded files are checked against the required dataset structure.
  • Evaluation before deployment: You can inspect training results and compare model quality before deploying.
  • Runtime monitoring: In the Runtime tab, you can monitor request volume, latency, and error rates for deployed models.

​
How do I refresh a model?

For datasets you upload yourself, go to the Data tab and click Add Files to provide updated data. Synced Topsort datasets update in real time. Updating a dataset and updating a trained model are separate steps. Retrain the model using the refreshed datasets, review its performance, and deploy the updated model when ready.

​
Is my data isolated from other retailers?

Yes. Your uploaded data and fine-tuned models are completely isolated per retailer and are not shared with other retailers.

​
What are the retention and deletion policies?

Retention and deletion policies can be configured based on your requirements. Discuss these requirements with your Topsort representative during setup.

​
Can I bring my own model?

Yes. T-Brain can host and deploy your own model. You can also use T-Brain embeddings as inputs to models you build yourself.

​
How do I connect a model to my store?

Your application calls a deployed model through an API and uses its response in experiences such as search or recommendations. API documentation is available on request, including authentication, request and response formats, latency, and rate limits. Talk to a Topsort sales representative to get started.