Relevance, in Practice.
What happens when you point a semantic reranker at an industry that's been misspelling itself for 40 years.
The Vocabulary Gap
Keyword search matches tokens. These share almost no tokens. Traditional search returns zero results or wrong results. The industry solves this with trained staff memorizing abbreviations. That doesn't scale.
I spent 14 years inside this industry at a multi-location auto parts distributor. I lived this problem daily: sales staff translating customer language into catalog codes, wrong parts shipped because of abbreviation confusion, returns eating margins. When ZeroEntropy's co-founder reached out about a Solutions Engineer role, I did what I always do — built the integration to see how the technology actually performs against a problem I know deeply.
Two Hats
Solutions Engineer integration meets domain expertise from 14 years inside aftermarket parts distribution.
What I built
- →Read the docs, installed the SDK
- →Built the adapter (lib/zeroentropy.ts)
- →Designed the reranker instruction template for automotive catalog abbreviations
- →Handled error cases, timeouts, missing keys
- →Built the before/after comparison tooling
- →Instrumented latency and score capture
- →Time to first working integration: 4 hours
export async function rerankWithHandling(
options: RerankOptions,
timeoutMs = 15000
): Promise<RerankOutput> {
const response = await zclient.models.rerank({
model: "zerank-2",
query: options.query,
documents: options.documents,
top_n: options.topN,
});
return {
results: response.results.map((r) => ({
index: r.index,
relevanceScore: r.relevance_score,
})),
e2eLatency: response.e2e_latency,
};
}export const AUTOMOTIVE_RERANK_INSTRUCTION = `You are reranking automotive aftermarket catalog parts. Catalog entries use industry abbreviations (RH/LH, BSM, HTD, W/O CAM, W/ CAM, PWR FLD). Match customer intent to catalog terminology — not literal token overlap.`;
Parts Advisor
A complete conversational flow — pre-set with synthetic data showing how semantic reranking and attribute-driven clarification work together.
This is a reference walkthrough. Connect a ZeroEntropy API key to run live queries against the synthetic catalog.
Baseline vs ZeroEntropy
Side-by-side comparison of keyword baseline retrieval and ZeroEntropy reranked results. Live scores appear when API is connected.
What ZeroEntropy Did vs What the Application Did
ZeroEntropy makes the candidate set more relevant. The application uses that better candidate set to conduct a smarter conversation.
What ZeroEntropy Did
- →Compared customer intent against catalog abbreviations
- →Bridged consumer language and catalog terminology
- →Reordered candidates by semantic relevance
- →Returned relevance scores
What the Application Did
- →Enforced hard fitment constraints (year, make, model, side)
- →Compared structured product attributes
- →Selected clarification questions based on attribute entropy
- →Applied customer answers to narrow candidates
- →Validated final compatibility
Ask My Work
A resume lists projects. This lets you ask what they prove.
Ask about my experience →Pipeline
How customer language flows through retrieval, reranking, and conversational narrowing.
Click a pipeline step to see what happens at that stage. The animation shows data flowing downward through the retrieval and reranking pipeline.