AN Alpesh Nakrani
SolutionsBlogBooksPraiseAbout Work with me ↗
eCommerce · Growth

Semantic search and a shopping assistant that end the dead-end query.

Understand natural-language shopping intent, retrieve eligible products, explain fit, and guide shoppers to the right PDP or cart without fabricating product facts.

◆ human-gateda person approves every consequential action
$15,500
fixed-scope pilot
Enterprise / regulated launch
launch posture
High current buying momentum
market signal
semantic-search-and-shopping-assistant
// parse intent and constraints
input: Query and session context
step: retrieve/rerank products
citations: [ source ✓ ]   confidence: 0.93
HUMAN GATEawaiting review →

Nothing is finalized until a human approves it.

Built for
The buyer
VP eCommerce, Chief Digital Officer, CMO
The champion
Search/Product Discovery Director, eCommerce Product manager, Merchandising lead
Day-to-day users
Shoppers, merchandisers, site-search analysts, customer service

Designed, built, and evaluated by Alpesh Nakrani, VP of Growth at ViitorCloud, 14 years shipping software, writing on AI-Native engineering and evaluation.

Evals-first
built in from day one
Human-gated
judgment stays with you
The problem

Where the time and money actually go.

Keyword search fails on conversational needs, attributes are incomplete, zero-result queries hide demand, and shoppers bounce when they cannot translate a need into filters.

Who feels it

Shoppers, merchandisers, site-search analysts, customer service

Trigger to act: Search exit or zero-result rates are high, catalog complexity is growing, agentic-shopping traffic is emerging, or the brand needs differentiated discovery.

Outcome & ROI

The result you can model before you sign.

Illustrative, replace with your data
$3,375
per month, illustrative

Illustrative only: $3 million monthly search-attributed GMV × 0.25% incremental conversion × 45% contribution margin = $3,375 monthly contribution. Run randomized experiments before scaling.

The outcome, plainly: Understand natural-language shopping intent, retrieve eligible products, explain fit, and guide shoppers to the right PDP or cart without fabricating product facts.

Search conversion
Zero-result rate
Search exit
How it works

Inputs in. A cited, review-ready result out. Your expert decides.

A Semantic search + shopping copilot. Every material fact is grounded in an allowed source and returned with its identifier, with no invented data.

01
Parse intent and constraints
02
Retrieve/rerank products
03
Apply inventory and merchandising rules
04
Explain recommendations with source attributes
05
Ask clarifying question
06
Add approved item to cart
Reference architecturegrounded · human-in-the-loop · fully auditable
Source systems · scoped access
Commerce platform
Product catalog/PIM
Search index
Inventory/pricing
Reviews
Grounded reasoning core
Retrieve & extract
grounded on your sources, returns citations
Reason & draft
Hybrid lexical + vector retrieval with a reranker; Gemini 2.5 Flash or GPT-5.4 mini
Human gateMerchandisers control ranking policies and exclusions; safety-sensitive or regulated recommendations escalate or use approved deterministic rules.
Action · only after approval
Add approved item to cart
Audit trace
sources, rules, confidence, reviewer
Tenant isolation
minimum data, never cross-tenant
Evaluation suite
baselined pre-launch, watched after
Observability
cost, latency & drift telemetry
Model strategy

Hybrid lexical + vector retrieval with a reranker; Gemini 2.5 Flash or GPT-5.4 mini for conversational answers; a stronger model is invoked only for complex comparison or policy questions. OpenSearch/Elasticsearch or Algolia; embedding service; catalog policy filter.

Inputs
  • Query and session context
  • Product attributes
  • Inventory
  • Price
  • Compatibility
  • Reviews
The human gate

AI-Native, not autonomous. Judgment stays with your people.

The machine does the work; the human’s role narrows to the one thing that matters, judgment. That constraint is what makes it safe to deploy.

Non-negotiable human gate

Merchandisers control ranking policies and exclusions; safety-sensitive or regulated recommendations escalate or use approved deterministic rules.

What it will never do
No fabricated compatibility or availability
No undisclosed paid ranking
No autonomous regulated-product advice
No promise that conversational search always improves conversion
The scorecard

A scorecard, not a demo. We baseline what breaks in production.

Every deployment ships with an evaluation suite. These are the numbers we baseline before launch and monitor after.

Primary
Search conversion
Zero-result rate
Search exit
Add-to-cart
Revenue per search
Irrelevant-result rate
Constraint violation
Recommendation citation accuracy
Why this, not that

The category is crowded. Most of it isn’t built for your workflow.

The alternatives
ConstructorAlgoliaBloomreachNostoKlevuCoveoGoogle Cloud Retail SearchShopify Search & Discovery
This implementation

Fixed-scope, tuned to your systems and rules, grounded in your data, with the human gate and audit trail built in from day one. A price you own, not a subscription you rent.

✓ Fixed price, not a seat subscription ✓ Grounded in your data & rules ✓ Human approval on consequential actions ✓ Auditable decision trace
Systems & integrations

It plugs into the stack you already run.

No rip-and-replace. Access is scoped to the minimum data necessary, isolated per tenant, and fully logged.

Commerce platformProduct catalog/PIMSearch index (Algolia, OpenSearch, Elasticsearch)Inventory/pricingReviewsAnalyticsConsented personalization
Pricing

Transparent by design. The build price buys the workflow and the proof.

A fixed implementation fee plus a monthly bill that scales with volume and governance. No hidden seats.

Pilot
$15,500
one-time · bounded proof of value
  • One process / scope
  • Live workflow on your data
  • Baseline evaluation suite
  • Measured vs. current process
Most chosen
Production
$36,000
one-time · full deployment
  • Full scope & integration
  • Human-review UI & audit trail
  • Write-back to your systems
  • Production evals & monitoring
Enterprise
$58,000
one-time · multi-entity / regulated
  • Multi-facility rollout
  • Advanced security & compliance
  • Custom control & escalation
  • Dedicated evaluation program
Monthly operating cost

100,000 to 1,000,000 catalog/search events per month plus embeddings, reranking and model answers.

$850–6,800
usage (models, OCR, vector, storage)
$2,200/mo
managed evaluation & monitoring

Planning assumptions, not vendor quotations. Your commerce platform and other platform licenses are separate and owned by you. Figures confirmed during scoping.

On working with Alpesh
“His vast knowledge of technologies and a natural problem-solving mindset consistently lead us through complex challenges with clarity and confidence.”
AM
Adil Multani
Senior Backend Developer
Why this is safe to try
01Baseline first. We measure your current numbers before we build anything.
02Fixed scope, fixed price. One process in the pilot. No open-ended engagement.
03Expand only if the scorecard earns it. You see the measured result before committing to production.
04Your people stay in control. The human gate means nothing consequential happens without a human’s approval.
FAQ

Questions serious buyers ask.

Does the AI act on its own?

No. Merchandisers control ranking policies and exclusions; safety-sensitive or regulated recommendations escalate or use approved deterministic rules. The system drafts and recommends; a human approves every consequential action. Explicitly excluded: no fabricated compatibility or availability; no undisclosed paid ranking; no autonomous regulated-product advice; no promise that conversational search always improves conversion.

How do you stop it inventing facts?

Every material claim is grounded in an allowed source record and returned with its source identifier. The system separates observed facts, model inference, and missing information, and routes to a human whenever confidence is low, evidence conflicts, or an adverse outcome is possible.

What does it cost to run each month?

A usage bill of roughly $850 to $6,800 per month (100,000 to 1,000,000 catalog/search events per month plus embeddings, reranking and model answers), plus a $2,200 per month managed retainer for evaluation, monitoring and maintenance. Your existing platform licenses are separate and already yours. Exact figures are confirmed during scoping.

Do we need a ChatGPT or Claude subscription?

No consumer ChatGPT or Claude subscription is required for the production workflow. The client needs an approved API/cloud billing account. Workspace seats are optional for internal prototyping and administrator access.

How is this different from Constructor?

Tools like Constructor, Algolia, and Bloomreach are broad platforms you adapt to. This is a fixed-scope implementation tuned to your systems and rules, grounded in your data, with the human gate and audit trail built in, and a transparent price instead of a seat subscription.

How long until it is live, and how do we prove it works?

We baseline “Search conversion” first, then measure against that baseline. You see the scorecard before expanding scope: the evaluation suite ships with the system, not as an afterthought.

Book a scoping call

Bring your real numbers. Leave with a fixed-scope plan.

A 30-minute engineering-led working session, no slideware. You leave with a sized opportunity estimate, a fixed-scope pilot plan, and the integration & human-review path mapped.

VP of Growth at ViitorCloud · senior delivery owner confirmed before paid work

Ask AI about Semantic Search & Shopping Assistant