Combine precision and meaning without invisible ranking.
Hybrid retrieval takes product candidates from multiple search methods and brings them together in one result list. Exact words and codes remain strong, while semantic matching helps with natural language. The value is in the controlled combination.
Why multiple retrieval methods?
No search technique is strongest everywhere. Lexical retrieval is accurate with model codes, brands and specific terms. Semantic retrieval finds meaningfully related products in applications and long descriptions. Hybrid retrieval uses both qualities without making one method blindly leading.
The system first collects candidates by source. After that, duplicates are merged, scores made similar, hard filters applied and the remaining products are rearranged again.
Candidates from different sources
- Lexical candidates based on terms, fields, synonyms and exact values.
- Semantic candidates on the basis of meaning agreement.
- Category or navigation candidates when the query clearly points to an assortment component.
- Business candidates which receive extra attention through controlled merchandising rules.
Each source provides a limited candidate set. There is no need to process the entire catalogue in every next step. That keeps the search route faster and more explainable.
Scores are not directly comparable
A lexical score and vector distance do not mean the same thing. They cannot therefore be added together without normalization. Scorefusion uses an explicit method to combine sources, for example, based on position, calibrated score, or query type.
The approach chosen should remain stable with catalogue growth and changes in one source. Otherwise, a technical score shift can unexpectedly change the entire result order.
Deduplication and variants
The same product can come from multiple sources. Deduplication prevents double tiles and combines the available signals. With product variants, an additional choice is needed: do you show each variant, group on main product or do you choose the best suitable variant for the search question?
That decision depends on assortment and user experience. A size or color question can require a specific variant, while a wide category search becomes clearer with grouped products.
Filters before final ranking
Hard conditions such as size, stock, country, customer rights or compatibility should not disappear in scorefusion. They are used as filters or constraints. A candidate who seems good but should not be delivered does not belong in the final list.
After that, a reranker can combine additional signals: exact field matches, semantic agreement, product quality, availability and recorded business rules. Each factor has a clear role and boundary.
Query-dependent priorities
With a SKU query, the lexical source is given almost all the space. With an application question, semantic retrieval can propose more candidates. With a brand plus attribute, the brand and attribute must remain protected, while meaning helps to find relevant product types.
Intent detection and confidence control these relationships. Low security can lead to a wider combination or a safe lexical fallback, rather than a hard semantic interpretation.
Practical example
For “light waterproof jacket for commuting bicycles” lexical retrieval finds products with words such as waterproof and jacket. Semantic retrieval can recognize shell jackets that are suitable for bicycles, even when “commuting” is missing. Filters monitor size and stock. The reranker raises products that explicitly support multiple desired properties.
A fashionable light jacket with no waterproof property should not only appear at the top by semantic likeness. Product data and constraints remain decisive for factual suitability.
Measure and manage
Evaluate by query type which source delivers relevant candidates and where the final order deteriorates. Monitor overlap, zero-result searches, latency and unexpected category shifts. Protect golden queries with regression tests.
Findoviq makes hybrid retrieval manageable by explicitly maintaining the combination of lexical, semantic, filters and business rules. This creates a broader search experience without the disappearance of accuracy and control.
Discuss your search questions
Do you want to know how this approach fits your assortment, product data and customer behavior? Together, we look at which query types have priority and where exact, semantic and business signals need to complement each other.