Choosing a hybrid retrieval approach: why Cartflow opted for model + rules search in 2026

AI · 6 min read

Cartflow built its initial search stack around large reranker models that used vector similarity and on-the-fly reranking to produce personalized results. While relevance improved in many queries, merchandisers complained about unpredictability and the lack of reliable control for promotions and safety rules. The product team faced a choice: double down on models for better personalization or add deterministic rules for business constraints.

After validating with merch and legal stakeholders, Cartflow designed a hybrid pipeline: the retrieval stage provides a ranked candidate set using embeddings, the ranking stage runs a light supervised model, and a rules layer enforces hard constraints and prioritizes promoted SKUs. Importantly, rules are explainable and surfaced in the admin UI so merch teams can preview impacts before publishing.

From a UX perspective, the hybrid approach reduced surprising placements and made result diversity easier to guarantee for seasonal campaigns. A three-month experiment showed conversion rate stability across promotions (+3% overall) and a 40% reduction in support tickets about “wrong” search results. The team also implemented a “why this” affordance in the product to explain when rules altered ranking, improving transparency for power users.

The lesson for startup product design is that pure ML often needs governance layers to match business realities. Hybrid pipelines let teams preserve model-driven personalization while adding deterministic control and explainability—an increasingly common pattern as ML moves from research to production in consumer-facing products.