Platform · 2025 – 2026
Open search & recommendation platform
A fully open-source design for search, recommendations and an “Ask AI” assistant on top of Magento, Vue Storefront and Strapi.
Hosted search and personalisation products work, but the data and the ranking logic belong to the vendor. This design keeps Magento as the source of truth for commerce and Strapi for content, and adds a first-party platform beside them that the company owns end to end.
Shape
- Backbone. Kafka carries catalogue changes from a Magento observer module, Strapi publish webhooks and first-party behavioural events from the storefront. GA4 arrives nightly through its BigQuery export.
- Stores. PostgreSQL holds the canonical catalogue mirror, ClickHouse the events and features, OpenSearch keyword search and facets, Qdrant the text, image, item and user vectors, Valkey the online features, MinIO the images and model artefacts.
- Models. Airflow trains ALS, item2vec and two-tower retrieval with a LightGBM ranker, tracked in MLflow and served through ONNX.
- Search. Hybrid BM25 and vector retrieval, search by image with OpenCLIP, and a retrieval-augmented “Ask AI” over the catalogue and Strapi content, served by vLLM.
- Language. Multilingual embeddings (
bge-m3,multilingual-e5) because queries mix Kurdish, Arabic and English.
Architecture and phased plan are complete; implementation is staged behind hosted services. In 2026 production search moved to Wizzy and then to Google Vertex AI Search, and a vector-database evaluation for retrieval, recommendations, image search and agents settled on Weaviate.