Skip to main content
RET-01 · Retail & E-commerce

Conversational Shopping Copilot

Outcome​

Let a shopper describe a need in natural language and receive grounded product recommendations with live price, promotion, and fulfillment promise, while the basket and consent state stay authoritative in Cosmos DB.

Portfolio attributeValue
Primary personaCustomer
Primary aggregateShoppingSession
Cosmos DB containers3
Grounding datacatalog, inventory, prices, promotions, reviews
Development estimateUSD 60-220 monthly dev estimate
Runtime optionsPython FastAPI and C# ASP.NET Core

Productivity hypothesis​

These measures reflect the workflow, evidence, decision boundary, and industry risk for this use case.

Impact typeMeasureConservative internal assumptionCustomer measurement
ProductivityQuery-to-satisfactory-answer time12-22% reduction(baseline median query-to-answer minutes - assisted median minutes) / baseline median minutes * 100
Quality or riskAccepted answers5-10 percentage-point improvementassisted percentage of answers accepted as satisfactory - baseline percentage

Important: Conservative internal pilot assumptions only; not public benchmarks, commitments, or expected customer results. Validate against a comparable baseline or matched control cohort with mandatory human review.

Rationale: Conversational retrieval can shorten product discovery, but stale price or availability data would invalidate the result.

Pilot guardrail: Use current catalog, inventory, price, promotion, and review data and disclose unavailable information.

Customer validation plan​

  1. Baseline the current workflow with representative users and tasks.
  2. Run the accelerator on the same task mix with human review retained.
  3. Measure query-to-satisfactory-answer time and accepted answers with the documented formulas.
  4. Replace internal assumptions with customer-observed results before a business case.

Why Azure Cosmos DB​

This AI solution needs an operational data plane, not only a vector index:

  • Hybrid vector and full-text retrieval combines semantic recall with exact identifiers, codes, names, and domain language.
  • Hierarchical partition keys and metadata predicates keep tenant and domain boundaries in the normal query path.
  • Operational JSON, source chunks, workflow and agent state, feedback, and audit records remain together for low-latency reads and governed writes.
  • Change feed reacts to source and operational updates without polling.
  • Autoscale throughput, configurable consistency, and multi-region distribution support variable demand and enterprise availability.

Domain workflow​

Domain workflow

StageActorInput → outputApproval
1. Capture shopping intent and constraintsShopperNatural-language need with budget and occasion → Structured intent frame on the ShoppingSessionPolicy controlled
2. Retrieve candidate SKUs with vector plus attribute filterRetrieval agentIntent frame and preference profile → Ranked ProductCard candidate setPolicy controlled
3. Validate price, promotion, and availabilityCommerce validation serviceCandidate SKU list and postal code → Priced, promisable shortlistPolicy controlled
4. Present cited recommendation and fit rationaleShopping copilotPriced shortlist with review digests → Comparison answer with source citationsPolicy controlled
5. Confirm basket update or checkout handoffShopperProposed cart mutation → Committed cart lines and order handoff tokenHuman required

Technical architecture​

Technical architecture

Cosmos DB data model​

ContainerHierarchical partition keyDocumentsSearch paths
shoppingSessions/tenantId, /shopperIdShoppingSession, ConversationTurn, ShopperPreferenceProfile/utteranceVectorEmbedding, /affinityVectorEmbedding, /utteranceVector, /affinityVector
merchandisingCatalog/tenantId, /productCategoryIdProductCard, PromotionOffer/descriptionVectorEmbedding, /reviewDigestVectorEmbedding, /offerTextVectorEmbedding, /descriptionVector, /reviewDigestVector, /offerTextVector
inventoryPromise/tenantId, /skuIdInventorySignalOperational fields only

Query contract​

QueryContainerScopePurpose
Active session replayshoppingSessionssingle-partitionRehydrate the current basket and the last turns for continuity across channels.
Promisable candidate lookupinventoryPromisecross-partition-boundedKnowledge-container query against inventoryPromise (bounded fan-out across the /tenantId + /skuId partitions named by the shortlist): filter ranked candidates to SKUs with available-to-promise at the shopper's preferred fulfilment nodes, as a bounded fan-out across the at most fifty shortlisted SKU partitions.
Eligible offer resolutionmerchandisingCatalogcross-partition-boundedKnowledge-container query against merchandisingCatalog (bounded fan-out across the /tenantId + /productCategoryId partitions named by the shortlist): return only offers currently in effect and stackable with the basket state, fanned out across the handful of product categories present in the basket.
Abandoned basket recovery scanshoppingSessionscross-partition-boundedBounded cross-partition control-plane query: find sessions with value still in the cart that stalled, for follow-up messaging. It cannot carry the second partition key because recovery targets are unknown until every shopper partition is examined, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by the ShoppingSession type inside a rolling 48-hour updatedUtc window, executed with MaxItemCount paging and continuation tokens.

Python and C# services route each document type to its declared container and hierarchical partition key. ETag-based writes protect mutable aggregates, while append-only records and bounded cross-partition queries preserve scale.

Azure deployment​

Azure deployment

Clone the source repository, authenticate with Azure, and run:

Set-Location accelerators/retail-e-commerce/conversational-shopping-copilot
az login
./Deploy.ps1 -Runtime python -Location eastus2

Change python to csharp to deploy the .NET implementation.

Explore the implementation​

Return to all accelerators