Architecture & Tech Stack

What's actually deployed, what the LLM is, where BC is on the wire, and the production-path pieces that are documented but not yet wired. Verified against deployed code on 2026-05-12. Per-slice Reality today in the Strategy panel (s) cites specifics; this page is the common stack.

Deployed surfaces

All six are live on the BigCommerce Testing Cloudflare account (5e70ef24…), project-namespaced as paradigm-*. The SE walks Andy through the first four; the last two support the build team.

Surface	URL	Stack
Prototype	`paradigm-prototype.pages.dev`	SvelteKit 2.57 + adapter-cloudflare 7 → CF Pages
Storefront	`paradigm-storefront.bigcommerce-testing-7727.workers.dev`	SvelteKit + adapter-cloudflare → CF Worker
Admin	`paradigm-admin.pages.dev`	Vite + React 18 + BigDesign 2.4 → CF Pages
API	`paradigm-api.bigcommerce-testing-7727.workers.dev`	Plain CF Worker (no framework) routing by URL prefix
Hive MCP	`paradigm-hive-mcp.bigcommerce-testing-7727.workers.dev`	CF Worker + D1 + Phase-1 token broker (per ADR-0007)
Hive Dashboard	`paradigm-hive-dashboard.pages.dev`	Static SPA on CF Pages

Request path — Live LLM (MM1.1, MM2.1, MM2.2, EE2.1)

Four of the thirteen capabilities exercise the LLM on the wire today. They all follow the same path through apps/api/src/lib/ai.ts:runChat:

Browser (slice page)
  │  form POST
  ▼
SvelteKit +page.server.ts            // runs in storefront or prototype CF Worker
  │  apiFetch('/mm1/chat', body)
  ▼
paradigm-api Worker                  // apps/api/src/scenarios/mm1.ts:handleChat
  │  runChat(env, messages, opts)
  ▼
apps/api/src/lib/ai.ts               // KV cache check first
  │   ┌──── hit?  → return cached text (+ trace 'cache_hit' to D1)
  │   └──── miss
  ▼
env.AI.run('@cf/meta/llama-3.1-8b-instruct', { messages, max_tokens, temperature })
  │  // Workers AI binding — runs on CF AI GPU pool
  ▼
response.response (string)
  │  extractJson<T>() — strips ```json fences, parses JSON object
  │  recordTrace() — D1 agent_traces table
  │  env.CACHE.put(key, text, ttl) — KV with TTL
  ▼
JSON back to SvelteKit page → DOM

Latency envelope: KV cache hit ≈ 10–50ms · LLM call (miss) ≈ 800–2400ms. Visible to the user via meta.model_latency_ms in the response JSON.

Worker bindings

From apps/api/wrangler.toml + apps/api/src/types.ts:Env.

Binding	Kind	Resource	Used for
AI	Workers AI	CF AI binding	All LLM calls via `env.AI.run(model, …)`
DB	D1	`paradigm-api-d1`	Scenario state (reorder_predictions, atrisk_signals, budgets, agent_traces, etc.)
CACHE	KV	`paradigm-cache`	Prompt-response cache · rate-limit counters · session bridge
UPLOADS	R2	`paradigm-uploads`	Documented for EE2.1 BOM photo input · not used in deployed text-parse demo
VECTORIZE	Vectorize	`paradigm-rag`	Documented for MM2.2 + EE2.1 RAG · deferred (CF token Vectorize Admin scope blocker)
EVENTS_QUEUE	Queue	`paradigm-events`	Producer + consumer for cross-scenario events

LLM call map — where the AI actually runs

Four runChat() callsites across the 13 capabilities. Other scenarios either heuristic-only (no LLM today) or stub.

Capability	File · line	Prompt cache key	TTL
MM1.1	`apps/api/src/scenarios/mm1.ts:50`	`mm1.1:<hash(msg)>`	1h
MM2.1	`apps/api/src/scenarios/mm2.ts:47`	`mm2.1:<sku>:<segment>`	24h
MM2.2	`apps/api/src/scenarios/mm2.ts:94`	`mm2.2:<anchorSku>`	24h
EE2.1	`apps/api/src/scenarios/ee2.ts:42`	`ee2.1:<hash(desc)>`	1h

Default model: @cf/meta/llama-3.1-8b-instruct (8B Llama 3.1 Instruct, Workers AI tier). Per-call model override is supported but not used today. All scenarios use the same model.

BC integration points — where BigCommerce is on the wire

Two BC API clients in apps/api/src/lib/. Three capabilities call them in the deployed code today.

bc-client.ts — BC core REST

Base: api.bigcommerce.com/stores/cdfqf9k6zf/v3/… · Auth: X-Auth-Token · Token: BIGCOMMERCE_ACCESS_TOKEN secret

Used by: nothing on the wire today (catalog queries documented but not deployed). Wires the future-state for MM1.1, EE2.1, MM2.1 catalog reads.

b2b-client.ts — B2B Edition REST

Base: api.bigcommerce.com/b2b/management/… · Auth: X-Auth-Token (legacy authToken deprecated 2025-09-30) · Token: BIGCOMMERCE_B2B_ACCESS_TOKEN secret (split per 2026-05-09)

Used by: EE1.1 registerPartner (POST /companies — creates real Company in sandbox) · EE1.4 unifiedOrders (GET /orders — pulls real B2B Edition orders) · MM1.3 handleCompliance (GET /companies/{id} for company name).

Sandbox store: cdfqf9k6zf. Seeded catalog: 6 real products (#2973-#2978) via scripts/seed-catalog.ts (commit 99d738d). Product IDs surfaced in MM1.1, EE2.1 LLM responses resolve to these.

What's NOT in the stack

Common questions an evaluator might ask. Truthful answers prevent overclaim damage.

No Anthropic Claude API. No OpenAI API. Pure Workers AI binding.
No tool / function calling. One-shot prompt → text → JSON-extract. The "tool-use over catalog" claim in some strategy panels is aspirational; the catalog is statically embedded in the system prompt as a hardcoded list.
No streaming. Single completion; the user sees latency as a "Parsing…" spinner, not token-by-token render.
No RAG / no Vectorize. Binding is commented out in wrangler.toml due to CF token Vectorize Admin scope. MM2.2 bundling + EE2.1 BOM matching documented to use RAG; deferred until the scope is granted.
No vision model for EE2.1 today. The "field photo → SKU" capability (Klevu's territory) is text-input only.
No BC catalog API on the wire for MM1.1 / EE2.1 / MM2.1. Productions reads are documented; deployed code uses hardcoded MOCK_PRODUCTS against the seeded sandbox SKU IDs.
No real payment processing for EE1.2. Settlement math demos the BC checkout integration surface; Stripe Connect / Adyen MarketPay / HyperWallet are the production answer.
No PROS pricing on the wire (ADR-0006 locked stub). MCP procurement handshake (EE2.3) — server is deployed, live Claude→MCP demo requires Claude Desktop config on the demo machine.

Hive coordination (separate stack)

The Hive layer (ADR-0003) is a multi-agent coordination + secret broker layer used by the build team. It's not part of the demo product surface. Listed here because Andy may ask "what's the Hive thing?"

What it is: A CF Worker (paradigm-hive-mcp) + D1 (paradigm-hive-d1) + dashboard. Cloned from bc-subscriptions/.hive.

What it does: Coordinates work across agent sessions (task claims, locks, decisions). Acts as a secret broker (per ADR-0007) so agents request short-lived BC API tokens scoped to their role rather than holding long-lived credentials.

What it's NOT: Customer-facing. End-users (buyers, merchants, Andy) never interact with Hive.

Reference docs

docs/SE-HANDOFF.md — SE launch page with 13-capability map
docs/content/02-feasibility.md §"Per-capability reality check" — truth column per capability
docs/decisions/0001-0007 — locked ADRs (storefront stack, infra, Hive, PROS stub, token broker, etc.)
apps/api/src/lib/ai.ts — Workers AI helper (cache + trace)
apps/api/src/lib/bc-client.ts · b2b-client.ts — BC core + B2B Edition clients
apps/api/wrangler.toml — Worker bindings + cron schedule