Which search method finds the right product most often?
Four search methods compared fairly on ten hand-checked questions.
A working sample we built to show how we would solve a problem like this. It is not a delivered project.
In plain terms
Four ways of searching a product catalogue were compared to see which finds the right item most often, using ten questions whose correct answers were chosen by hand beforehand. Combining keyword search with meaning-based search and then re-ranking the top results did best. Product details written in Chinese and English were also turned into one consistent format.
- All four methods were tested on the same questions.
- The results are measured, not claimed.
- Mixed-language product data became one consistent format.
What this does not claim: Only ten test questions, so the results show that the method works and not how it will score on a much larger catalogue.
Technical details
Four retrieval methods (BM25, dense-only, RRF hybrid, hybrid + re-rank) compared on a 10-query hand-labelled gold set, scored by Precision@3 and Recall@3. Also JSON-Schema-enforced structured extraction with per-field confidence, applied to Chinese-sourced product listings and normalised to the English catalogue schema.
- PostgreSQL 18 + pgvector (HNSW)
- BM25 (rank_bm25)
- Reciprocal Rank Fusion
- Cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2)
- JSON-Schema structured extraction
Need something like this?
Tell us the problem in your own words. We reply with a plain, practical plan.
Discuss a project like this