Startup Hybrid Search

This demo clones into a working hybrid search engine over a catalog of startups, letting a developer compare three retrieval modes on the same query and data: semantic search that embeds the query with a dense model and ranks by vector similarity, keyword search that ranks with BM25 over a sparse vector with server-side IDF, and a hybrid mode, the default, that runs both and fuses them with reciprocal rank fusion.

Qdrant performs the retrieval end to end, holding a named dense vector for semantic search and a sparse BM25 vector for keyword search in the same collection. It runs fully locally against a Qdrant container with Docker and Python, embedding with fastembed, or can point at Qdrant Cloud so embedding happens server-side instead. A search API returns a stats object naming the active mode and its score type, since the three scales are not directly comparable.