AI Search and Matching for semi-structured data

Uses Mixture of Encoders with Qwen3-0.6B for product description and category encoding, numerical encoders applied to product ratings, rating counts and prices. Can also generate query-specific filter predicates against materials, colors and style properties. Configured with GPT-4o for the query understanding module.
No re-ranking or metadata boosting.

Azure AI Search with Semantic Re-ranker and the built-in query understanding functionality powered by OpenAI LLM API.
We implemented multiple configurations available in the Azure AI ecosystem and we took the best results provided by each - for details see link below.

Vertex AI Search with built-in Hybrid Search configured with gemini-embedding-001 dense embedding model, splade-v3 sparse embedding model and Vertex AI RankingAPI.
The Superlinked configuration performs significantly better despite gemini-embedding-001 outperforming Qwen3-0.6B by 3 percentage points on MTEB.

Use SearchRequest from the Vertex AI Discovery Engine with Boost, QueryExpansion and built-in re-ranking. Tested with and without query parsing using built-in Gemini feature and took the best results.

The indexed JSON objects were "stringified" and embedded with Qwen3-0.6. The same model is used to encode the queries. The single dense query vector is used to retrieve the relevant results, without re-ranking.