Skip to main content

Reference showcase — Atlas Store

Atlas Store is the canonical reference application for Viglet Turing ES — the "see it all wired together" artifact. It's a commerce-catalog copilot whose single job is to exercise the maximum surface of the platform: a living demo, an SDK-vs-real-backend integration test, and the answer to "what can Turing actually do?"

If you've read the feature pages one at a time, this is where they connect. Atlas Store is the through-line: a typed catalog feeds faceted + hybrid search, the same index grounds a RAG chat, the chat gains agent power (client tools, code interpreter, skills), and a handful of creative compositions only possible because Turing owns search and agents. This page walks that surface in the order a visitor experiences it, linking the deep page for each capability — making it a good map for the learning path.

It lives in the monorepo at frontend/apps/showcase and seeds its own data: a deterministic ~200-product typed catalog (price as CURRENCY, rating as FLOAT, reviews as INT, on-sale as BOOL, release date as DATE, multi-valued color/material/tags) that stresses the manifest, hybrid ranking, and facets.

Run it

# From the monorepo root
pnpm install
pnpm dev:showcase # vite dev server (proxy VITE_API_URL → a running Turing)

# Package as an import bundle
pnpm build:showcase
cd frontend/apps/showcase && npm run zip # → atlas-store.zip

Import atlas-store.zip from the Turing admin (Marketplace / Pages import) to provision the SN site, index the catalog, and serve the SPA as a search template.

Two fidelities

The offline Lucene seed is search-only (no embeddings) — chat and agent surfaces show honest "configure GenAI" notices rather than breaking. Import the bundle into a backend with an embedding LLM + AI agent configured and the chat, hybrid ranking, and agent features light up. Every agent-side feature is gated server-side; the client wiring is always present.


The walkthrough

1. Search — the typed catalog

The storefront is a faceted search built on the @viglet/turing-react-sdk search hooks: URL-synced facets, autocomplete, a "did you mean?" spell-check banner, category tabs, a sort dropdown wired to the site's custom sorts, a "you may also like" similar rail, and click-tracking on every result. Because the catalog is typed (a field manifest), CURRENCY sorts as money, DATE filters as a year range, and multi-valued tags facet with OR semantics. With an embedding model configured, hybrid RRF ranking + a reranker are a server-side site-config toggle — the UI is rank-mode agnostic. → Semantic Navigation, Search Engine

2. Ask Atlas — the RAG chat

A floating "Ask Atlas" copilot exercises the RAG chat half via useTuringChat: token streaming, cited sources as chips, option chips for the next turn, native multi-field forms (text/email/date/select widgets), a slots "Collected" strip, multimodal image upload with vision, a workspace panel of agent artifacts, and rich content dispatch — markdown, sandboxed html, and d2 diagrams (the diagram engine lazy-loaded so the main bundle stays small). → Chat, RAG

3. Agent power

The copilot is a real AI Agent, so it can act, not just answer:

  • Client toolsadd_to_cart / get_cart / clear_cart run against a real cart context shared with the header drawer, so the agent acts on the shopper's live basket. → Client Tools
  • Code interpreter — "total my cart" / "chart price history" runs Python over real data. → Tool Calling
  • Live tool-call activity — each tool invocation streams inline as it runs. → Client Tools → Live tool activity
  • A bundled returns/RMA skill — a self-contained Anthropic skill folder (SKILL.md + a reference doc + a pure-stdlib rma.py) you drop into the backend skillsPath. → Skills

4. Creative compositions

These pay off precisely because Turing owns both search and agents. They are shipped, opt-in features (default off) — each enabled per agent, demonstrated here wired together:

CompositionWhat it does
Answer-as-an-appThe agent renders a live comparison_table / spec_card / configurator from a real typed search, via generative UI
Co-browseThe agent drives the actual storefront the shopper sees (set query, toggle facets, sort, page)
Glass-box "Why #1?"The agent explains the top result's ranking with the real signals (BM25 / vector / RRF / rerank), via the explain_ranking tool
Ambient proactive copilotA toast offers help when an interaction signal crosses the agent's threshold (opt-in, throttled)
Agent→agent handoffA router agent delegates to a specialist via delegate_to_agent — visible for free in the tool-call activity
Skills that ship UIA skill folder can embed its own components, surfaced through the same generative-UI tool path
Embeddable action widgetA mock third-party page embeds the agent via the zero-dep vanilla @viglet/turing-sdk, so the agent acts on that site's cart — Turing as an action layer over any website
Opt-in and default-off

Each composition is gated by its own per-agent flag (answerAsAppEnabled, coBrowseEnabled, proactiveEnabled, agentHandoffEnabled, userMemoryEnabled, actionWidgetEnabled, …). They are real, tested capabilities — but off by default, so a stock deployment never changes until you opt one in. The showcase turns them on to demonstrate the full surface.

5. Voice, memory & ops

  • Voice — a mic button dictates the question (Web Speech STT) and auto-sends on a final utterance.
  • Search by image — a hero affordance hands off to the copilot's multimodal vision upload.
  • Cross-conversation memory — a "brain" panel shows what Atlas remembers about the visitor (keyed on a stable visitor id), with a "Forget me" GDPR delete. → Chat Memory
  • Ops tour (/ops) — a guided map of Chat Analytics (sentiment trajectory, tool-latency p95, SSE channels), agent eval + golden sets, self-tuning suggestions, and cost governance.
  • MCP exposuremcp/claude-desktop.json bridges Claude Desktop to Turing's MCP server, so the Atlas catalog is drivable from Claude Desktop.

An in-app Guided Tour narrates each capability with "try it" deep-links into the copilot.


What it demonstrates (at a glance)

SurfaceCapabilities
Searchfaceted + URL-synced search, autocomplete, spell-check, category tabs, sort, similar/"more like this", click tracking; hybrid RRF + reranker as a server toggle
RAG chatstreaming, cited sources, option chips, native forms, slots, multimodal image upload, workspace panel, rich markdown/html/d2
Agent powercart client tools, code interpreter, live tool-call activity, answer-as-app widgets, a bundled returns/RMA skill
Creative compositionsco-browse, glass-box "Why #1?", proactive copilot, agent→agent handoff, skill UIs, an embeddable action widget
Voice / memory / opsvoice dictation, search by image, cross-conversation memory (GDPR delete), an /ops tour, MCP exposure for Claude Desktop

A full feature-by-feature reconciliation against the design matrix (✅ live / 🔌 backend-gated / 📄 documented) lives in COVERAGE.md.

Architecture notes

  • Built with @viglet/turing-react-sdk (search/chat hooks), @viglet/turing-react-ui (headless components, token-themed to the Atlas brand), and the zero-dep vanilla @viglet/turing-sdk (the embeddable action widget).
  • Agent-side features are gated server-side; the offline Lucene seed is search-only and the UI shows honest "configure GenAI" notices, so the demo never looks broken.
  • An end-to-end smoke test (TurShowcaseSmokeIT) boots the backend, imports the showcase ZIP into an embedded Lucene index, and asserts the REST contract the storefront depends on.

PageDescription
Getting StartedThe zero → first-agent learning path this showcase makes concrete
Semantic NavigationThe faceted search powering the storefront
RAGThe retrieval that grounds "Ask Atlas"
AI AgentsThe agent behind the copilot
Client ToolsCart tools, generative UI, the embeddable widget
React SDKThe hooks and components Atlas Store is built on