Search Engine

How scoring works

Each search result receives a final score computed from four weighted components:

• Vector — cosine similarity between the query embedding and the product embedding (BGE-M3 dense + sparse).
• Fuzzy — lexical similarity using rapidfuzz WRatio with Armenian suffix stemming.
• Expansion — bonus for products matching synonym / slang expansions (not the original query).
• Cat/Brand Boost — additive bonus when the query overlaps the product's category or brand name.

Formula: final = vector * W_vec + fuzzy * W_fuz + expansion * W_exp + boost * W_boost

Threshold parameters

• Text Score Threshold — minimum cosine similarity for text search. Increase to get fewer but more relevant results.
• Image Score Threshold — minimum SigLIP 2 similarity for image search. Filters out visually unrelated items.
• Hybrid Alpha — balance between text (1.0) and image (0.0) signals in hybrid search.
• Reranker Cut-off — minimum cross-encoder confidence after deep reranking. Removes low-relevance semantic matches.

NER category filtering

V2 text search uses Named Entity Recognition to detect brand and category names in the query. When a match is found, a hard Qdrant filter is applied (not just a boost). Only categories that contain at least one product are eligible for NER matching — parent grouping categories with no products are automatically excluded to prevent empty results.