BUILT WITH ZEROENTROPY
Relevance, in Practice.
What happens when you point a semantic reranker at an industry that's been misspelling itself for 40 years.
The Problem
The Vocabulary Gap
What the customer says
passenger mirror with the warning light
What the catalog says
MIRROR ASSY DOOR RH PWR HTD BSM W/O CAM
What the customer says
the mirror with a camera underneath
What the catalog says
MIRROR ASSY DOOR RH PWR HTD BSM W/ CAM PWR FLD
What the customer says
I need the right headlight
What the catalog says
HEADLAMP ASSY RH HALOGEN W/ LED DRL
What the customer says
espejo derecho con punto ciego
What the catalog says
MIRROR ASSY DOOR RH BSM
Keyword search matches tokens. These share almost no tokens. The industry solves this with trained staff memorizing abbreviations. That works until you have 30,000 SKUs, 100,000 fitments, and customers calling at 10pm when nobody's there to translate.
The Cost
What Bad Search Actually Costs
23%
of returns
Caused by wrong-part shipments from search errors.
$45
avg cost per return
Shipping, restocking, customer time, and lost trust.
3.2 min
translation time per call
Rep time spent translating customer language to catalog language.
These are industry-representative figures, not from a specific customer. A real engagement would start by measuring yours.
Before vs After
What ZeroEntropy Changes
Query
passenger mirror with the warning light
BASELINE RESULTS (keyword)
#1 MIRROR, INTERIOR, REARVIEW
0.34#2 MIRROR ASSY DOOR LH PWR
0.31#3 MIRROR ASSY DOOR RH PWR HTD
0.28ZeroEntropy results (zerank-2)
#1 MIRROR ASSY DOOR RH PWR HTD BSM W/O CAM
0.94
moved up
#2 MIRROR ASSY DOOR RH PWR HTD BSM W/ CAM
0.91
variant match
#3 MIRROR ASSY DOOR RH PWR HTD BSM W/ CAM PWR FLD
0.89
variant match
Baseline got the type wrong (interior vs exterior), side wrong (LH vs RH), and missed BSM. ZeroEntropy maps warning light to BSM and passenger to RH.
Captured from live RapidFleet search comparison (2026-06-30).
Live API integration is available with a ZeroEntropy API key.
The Conversation
From Search to Sale
Pre-set walkthrough for Query: “I need the passenger-side mirror with the warning light” · Vehicle: 2021 Ford F-150
Outside door mirror or inside windshield mirror?
Answer: Outside
Candidate reduction: 18 candidates -> 9
Final recommendation
MIRROR ASSY, DOOR, RH, PWR, HTD, BSM, W/ CAM, MAN FOLD
Best compatible catalog match — verified against confirmed attributes.
4 questions. 30 seconds. The right part. No training manual. No memorized abbreviation table. No call back from the warehouse.
The ROI
What This Means for the Business
31% -> 94%
Search accuracy
Baseline Top-1 vs ZeroEntropy Top-1
-68%
Wrong-part returns
Projected reduction from correct first-match
3.2m -> 0.8m
Rep time per call
Translation time eliminated
0% -> 100%
After-hours coverage
AI agent with good search works 24/7
These projections are modeled from reference evaluation and live comparison snapshots. A real engagement measures the customer's baseline first, then tracks improvement.
Two Hats
Why I Built This
The Solutions Engineer hat
I wanted to understand how fast a developer can integrate ZeroEntropy, where it fits in an existing pipeline, and what edge cases matter.
- Read docs, installed SDK
- Built adapter and instruction template
- Handled errors and fallbacks
- Added baseline vs reranked comparison UI
- Time to first working integration: under 1 hour
The domain expert hat
I spent 14 years translating between customer phrasing and catalog shorthand. That knowledge lives in instruction design, not docs.
- RH/LH mapping for US-market LHD vehicles
- BSM = blind spot = warning light = little triangle icon
- Ask visible features first, technical specs last
- Design for 30,000 SKUs and 100,000 fitments
Stakeholder View
Who Cares and Why
| Stakeholder | What success means |
|---|---|
| Customer | Find the right part fast |
| Sales Team | Less translation, fewer errors, more calls handled |
| Operations | Fewer returns, accurate fulfillment |
| Engineering | Clean integration, reliable infrastructure |
| Business Owner | Revenue, efficiency, customer satisfaction |
A successful implementation isn't just better search. It's better search that every stakeholder can measure.
Architecture
End-to-End Decision Pipeline
Selected node
Customer query
Natural language input, often shorthand-heavy or multilingual.
If I Had Another Week
Next Validation Layer
- Retrieval relevance scoring with real test set
- Benchmark query dataset (50+ queries)
- Search failure analysis and root cause breakdown
- Latency measurements by query type
- Voice interaction testing (Twilio + Bland)
- A/B comparison methodology for customer deployment
About the author
Philippe Chaunu — Miami, FL
I build operational software and applied AI where customer needs, messy data, and production reality meet. This project reflects how I learn new products: implement first, evaluate second, communicate third.
US & French dual citizen. Trilingual: English, French, Spanish.