WORKSHOP
Towards a Scalable Personalized Search Application for E-commerce
In this workshop, we’ll work on the building blocks of an E-commerce search application that deals with both explicit and implicit preferences (e.g., selected an attribute vs .previous clicks and cart adds), and also performs well. We’ll use Vespa.ai, an open-source AI search platform, to do:
-
Quick first-stage retrieval and ranking using BM25 keyword search
-
Dense tensor ANN for vector search (hybrid with the previous BM25). This also scales well; typically ANN is much faster than BM25 in Vespa.
-
Sparse tensor dot product for re-ranking based on user preferences
Participants will understand how to adapt these techniques to their own use cases and also get other ideas for further development (e.g., using ONNX models for re-ranking).