Free text in, 0–3 ranked matches out, each with a match %, a "why it matches you" tied to the user's own words, and a serving suggestion.
Multi-turn refinement in the same session ("something cheaper", "less smoky") - "cheaper" becomes a numeric price_level filter on retrieval.
Server-side checks on every pick: unknown ids, out-of-stock bottles, duplicates and low scores are dropped and logged; scores clamped, re-sorted, cut to N.
Off-topic input gets a polite steer and zero matches; distress gets a gentle answer and no alcohol push; prompt injection cannot produce non-inventory bottles or reveal the prompt.
Structured output validated with Pydantic, one retry, then a clean AGENT_ERROR. A failed turn is rolled back so the session (and the UI Retry) keeps working.
POST /chat with request validation, warm-up requests, and contract error codes for validation, throttling, timeouts and agent errors.
Stage throttling 5 rps / burst 10; Lambda timeout below the API Gateway 29 s limit.
Chat with example prompt chips, typing indicator, ranked cards (score bar, details, tasting notes, why, serving), "nothing in our cellar fits" state, error banner with Retry, New conversation, responsible-drinking footer.
Fixtures mode (make ui-fixtures) to work on the UI with no backend.
One stack, make deploy / make destroy; agent-only redeploys for prompt or model changes; make kb-sync for inventory changes.
One JSON log line per agent turn: input, retrieved ids, raw picks, dropped ids and reasons, tool calls, latency.
place_order tool exposed through an AgentCore Gateway Lambda target (IAM inbound auth from the runtime role), mock order Lambda returning a UUID4.
Two-turn confirmation enforced in code: the first order request only returns a pending_order; a Strands hook blocks place_order unless the previous turn proposed exactly the same items and quantities.
UI: per-card checkbox and quantity stepper, Order selected (n), Confirm / Cancel, order confirmation card.
Golden prompts 9–10 and unit tests for the order guard and order Lambda.
Built
Next