How it works

Three ways to rank the same query.

Every search runs keyword, semantic, and hybrid retrieval side by side — so the difference between matching strings and matching meaning is impossible to miss.

Retrieval modes

One query, three rankings

Each path has a real engine behind it — no hand-waving.

Keyword

BM25 · Elasticsearch

Ranks by literal term overlap across title, brand, category and description. Fast, transparent, and the honest strawman every semantic system has to beat.

Semantic

dense vector · pgvector

Embeds the query with bge-small and ranks by cosine similarity — so “affordable notebook” finds the budget Chromebook that shares no keywords.

Hybrid

reciprocal rank fusion

Fuses both rankings with RRF under a tunable α. It recovers the literal match and the intent match — usually the best relevance of the three.

Worked example

“cheap gaming laptop”, three ways

Keyword nails the literal match but misses the value picks. Semantic reads the intent but drops the exact model. Hybrid keeps both.

	Keyword	Semantic	HybridBEST
Top result	Gamer Pro 17	Budget Chromebook	Gamer Pro 17
Reads “cheap”?	No — literal terms only	Yes — as budget intent	Yes
Typical miss	Budget / value picks	The exact model asked for	—
Best for	Exact names & SKUs	Synonyms & paraphrases	General shopper queries

The request path

What happens on every search

Parse the query

Natural-language price filters are pulled out (“laptop under 300”) before retrieval runs.

Run three engines

Keyword, semantic, and hybrid execute in parallel on every search.

Fuse & rerank

RRF blends rankings; an optional cross-encoder refines the top candidates.

Score & diversify

MMR suppresses near-duplicates; NDCG/Recall/MRR are computed on labeled queries.

Watch the three paths race.

Run any query in the demo and compare keyword, semantic, and hybrid side by side.