S1. Natural-Language Semantic Search (AMWAY)

📅Published 2026-05-14·🔄Updated 2026-06-30·⏱3 min read

1. URL & Personas

• /semantic-search · P1 (ABO Field), P2 (Insights), P5 (Compliance)

2. User Stories

• P1: "Vegan protein + new ABO in their 30s + no active subscription" → instant ABO recommendations
• P5: "Review Nutrilite ad language + SKUs with minor risk" → regulatory guard list

3. Input UI

• Natural-language input box (multilingual support)
• Filters: category (Nutrition / Beauty / Home), brand, ABO level, country, price
• Persona switcher

4. Data Mix

Data	Source	Use
In-house SKUs	OpenSearch idx_product	BM25 + KNN
ABO profile	OpenSearch idx_abo	KNN
In-house reviews	OpenSearch idx_review	BM25 + KNN
External social posts	OpenSearch idx_social_trend (Reddit · Instagram · X)	KNN (multilingual)
Regulatory text	OpenSearch idx_regulation	BM25 (P5 only)
KG 1-hop	Neptune (ABO × Product × Brand × BU)	Graph expansion

5. Processing Pipeline

1. Input → Cohere embed-v4
2. OpenSearch in parallel: BM25 + KNN → top 100/100
3. RRF fusion → top 50
4. Cohere rerank-v3.5 (cohere.rerank-v3-5:0) → top 20
5. Neptune 1-hop expansion (ABO / SKU nodes)
6. Result cards + graph

6. Output UI

• Three columns: SKU cards / ABO · Customer cards / external review cards
• 1-hop KG (cytoscape.js)
• Cohort badge: 🟢 real / 🟡 synth / 🔵 external
• Result count + response time

7. Per-department Variations

Persona	Suggested Query
P1	"Diamond candidate ABO recommendations"
P2	"Emerging vegan trends matched to in-house SKUs"
P3	"Review embedding drift"
P4	"Subscription churn risk + similar ABOs"
P5	"SKUs with ad-language risk"

8. Guardrails

• PII masking (ABO and customer names and contacts)
• Block nutritional supplement and beauty marketing to suspected minors
• Cite external review sources

9. Demo Scenarios

1. "Vegan protein + 30s + Instagram +20%" → Nutrilite Plant Protein + reviews
2. "Diamond-level candidate ABOs" → ABO + tree (linked to S9-A)
3. "Artistry ad-language review" → SKU + regulatory risk markers