Hybrid search
Architecture and rollout of a hybrid search that combines semantic retrieval on embeddings with lexical search, for three of Switzerland’s leading e-commerce platforms.
- +10.7%
- Click-through rate
- +2.6%
- Conversion
- 3
- E-commerce platforms served
01
The idea
- Lexical search matches the words a customer types.
- Semantic retrieval on embeddings matches what they mean.
- Hybrid search combines both: exact queries like a product number still hit, vague ones still find the right products.
02
My role
- Led the architecture, from first design to deployment.
- Chose the vector search engine, weighing building against buying.
- Ran the alignment between data science and engineering.
- Key contributor to the search behind three of Switzerland’s leading e-commerce platforms.
03
What was hard
- Bringing a vector search engine in next to the existing lexical search.
- Integrating it concretely into the live search stack, not as a side experiment.
- Caching to keep it fast and cheap, both for live queries and in the import flow.
- The import flow: keeping product data embedded and indexed as the catalogue changes.
- The embedding service that turns products and queries into vectors.
- Hard constraints: costs had to stay low, latency couldn’t change meaningfully, and system reliability had to stay unaffected.
04
Outcome
- Proven in an A/B test against the existing search.
- +10.7% click-through rate.
- +2.6% conversion.
- Costs stayed low and latency didn’t change meaningfully.