Birdie, the AI Shopping Assistant

The shopper talks to a chat module in the storefront, which streams replies word by word and renders product cards, source links, and a permanent notice that the answers come from an AI assistant. The storefront's backend signs a short-lived statement of who the shopper is with a managed key service, and the assistant never holds the shopper's password or account credentials. A shopping chat service verifies that statement, holds the conversation in a managed key-value store, and runs the agent loop of model, tools, model. Before each turn it rewrites follow-up questions into a standalone search and passes only sufficiently relevant passages of store content to the model.

A dedicated tool server exposes the shopper tools over an open agent protocol, grouped by domain: product search and refinement, filter options, product details, comparison, related products, the category tree, cart, and shopping list. It is never exposed to the browser directly. Separate commerce services, one per capability, each hold their own least-privilege access to the commerce platform, and cart changes use the shopper's own session. Every product, price, and availability claim comes from a live call during the conversation, never from the model's general knowledge.

A content pipeline re-indexes pages whenever an editor publishes or unpublishes them, with a nightly full re-index as a backstop. A queued indexer splits pages into passages, creates embeddings on the same model platform that runs the agent, and stores them per language. A hybrid index combines vector similarity for paraphrases with keyword search for product names and model numbers. Content sources are pluggable behind one interface.

An independent guardrail checks every question and answer, blocking off-topic requests, prompt attacks, and harmful content, and masking or blocking personal data such as card numbers. Third-party text in product data is treated as data, never as instructions. Cart quantities are absolute and every change carries an idempotency key, so a retry cannot add a product twice. Checkout and payment always stay in the storefront's own checkout. The shopper-facing agent is kept apart from supersonic's merchant-facing copilot, each with its own commerce access, so it can never reach tools that change prices or show other customers' data.

Traces and metrics flow to the cloud provider's observability services, and conversations are recorded with personal data masked. The agent and the content search are evaluated on every change in continuous integration. All infrastructure is defined as code and runs in the provider's Frankfurt region.