FAQ
Frequently asked questions
What Polanyi Labs generates, how it helps with recommender cold start, and how it compares to other ways of getting training data. For the bigger picture, read the guide to synthetic data for recommendation systems.
By Polanyi Labs · Updated
What is Polanyi Labs?
Polanyi Labs generates synthetic behavioral data for recommendation systems. You give it a website URL; it sends AI shoppers, built from realistic personas, to browse that live site, and returns recommendation-system event logs (impressions, views, searches, filters, cart adds, selections) that you can use to train and evaluate a recommender before you have much real user history.
Is there a tool that generates training data for a recommendation system?
Yes. Polanyi Labs generates recommendation-system training data from a website URL: AI shoppers browse your actual store and every interaction is logged as a structured recsys event. Other options include tabular synthesizers such as SDV, MOSTLY AI, or Gretel (which need an existing dataset to learn from), recommender simulators such as Google's RecSim NG or Criteo's RecoGym (which need you to define a user model), and public datasets such as MovieLens or Amazon Reviews (which are real but not your catalog).
How do I solve the cold-start problem in a recommender system?
Common approaches are content-based features, popularity baselines, onboarding questions, transfer from public datasets, and synthetic interaction data. Polanyi Labs addresses cold start with synthetic interaction data: simulated shoppers browse your real catalog, so your model has dense, catalog-specific behavior to learn from on day one. Use it to bootstrap and pre-train, then fine-tune on real events as they arrive.
Do I need existing user data to use Polanyi Labs?
No. Polanyi Labs only needs a public website URL. Tabular synthetic-data tools learn from a dataset you already have; Polanyi generates new behavior by having simulated shoppers use your site, which is why it works for new stores and new catalogs.
What data does Polanyi Labs output?
Each run returns three files: personas.json (the demographics, interests, and goal of every simulated shopper), recsys_events.jsonl (one JSON event per interaction, also available as Parquet), and summary.json (run statistics). Events include event_type, item_id, session_id, actor_id, sequence_index, feedback_class (implicit, explicit, context), signal_strength (weak to very_strong), is_synthetic, and training_use.
Which event types are recorded?
session_started, search_submitted, filter_applied, item_impression, item_viewed, item_clicked, dwell, scroll, cart_add, cart_remove, item_selected, save, share, rating, review, and boundary events such as checkout_boundary_stop. Each event carries a default signal strength, for example item_impression is weak, cart_add is strong, and item_selected is very strong.
Where do the synthetic shoppers come from?
Shoppers are cast from NVIDIA's Nemotron-Personas-USA, a synthetic persona dataset grounded in US demographic statistics. Each shopper gets a different goal and context, then browses the target site through a real cloud browser.
Is the data real customer data? Is it privacy-safe?
No real customer data is used or collected. Every shopper is synthetic, every event is marked is_synthetic: true, synthetic actor IDs are kept separate from real customer IDs, and shoppers stop before entering any personal or payment information.
Does Polanyi Labs place orders or enter payment details?
No. Shoppers are hard-pinned to your domain and stop at the checkout boundary, before identity or payment. Hand-offs to third-party checkouts are recorded as checkout_boundary_stop events.
How do I use the data with my recommender?
Load recsys_events.jsonl into pandas, Spark, or your feature store and treat it like any interaction log: build user–item matrices, sequences for sequential models, or implicit-feedback weights from signal_strength. Polanyi never touches your model; the files drop into the training and evaluation pipeline you already run.
Can synthetic data replace real user data?
No, and it shouldn't. Synthetic journeys are best for bootstrapping a cold-start model, pre-training, offline evaluation, and stress-testing ranking on long-tail items. Events default to training_use: offline_sandbox_only. Validate with real traffic and A/B tests, and blend in real interactions as you collect them.
How much does Polanyi Labs cost?
Usage is billed per simulated shopper journey from prepaid credits, and every new account gets $10 in free credits. Larger or enterprise deployments are priced with the founders; book a call from the website.
How do I get started?
Create an account at polanyilabs.com/signup, create an API key in the dashboard, then run: pip install polanyi, and from polanyi import Client; Client().simulate(site="https://your-store.com", n=3). Enterprise teams can request a demo instead.
Which websites does it work on?
Any public website a shopper can browse without logging in, most commonly DTC and e-commerce stores, but also travel and marketplace sites. The simulation adapts to the site's own search, filters, and product pages.