Case study

Conversational Product Discovery

Semantic search and an advisory layer for shoppers who know the occasion but not the product — validated with a three-day prototype before any production spend.

Client type: Online retailer with a large gift-card catalogue (sample scenario)
Industry: E-commerce
Duration: 5 months

Sample contentSample engagement. Illustrative scenario used while the site is in build — not a named client project.

e commerce2026

TypeScript
Python
LLM integrations
RAG

01Challenge

Keyword search fails the largest hesitant segment. Shoppers arrive describing a person and an occasion, not a product, and the catalogue only responded to exact terms. Vague queries returned nothing and the session ended.

02Solution

A retrieval layer built on embeddings rather than keywords, an intent-comprehension step that tolerates typos, slang and emoji, and an advisory layer that explains why each suggestion fits the described situation.

03Business impact

A segment that previously bounced on empty search results was converted into browsing sessions, and merchandising gained visibility into what shoppers were asking for but could not find.

04Approach

01Three-day throwaway prototype to test whether intent matching worked at all
02Query log analysis to build an evaluation set of real, previously failing searches
03Embedding-based retrieval over catalogue and editorial content
04Advisory responses constrained to catalogue facts to prevent invention
05Progressive rollout behind a flag with side-by-side comparison against keyword search

Prototype first, infrastructure later

The riskiest assumption was whether intent matching would beat keyword search on real queries. A disposable prototype answered that in days. Production infrastructure was only built after the answer was yes.

Start a conversation

Conversational Product Discovery

Prototype first, infrastructure later

Services involved

More case studies

Similar problem on your side?Let's talk.