Semantic Search (Planned)

This use case is in development. Want to contribute? See the arcadedb-usecases repository.

Standalone vector search for e-commerce product discovery and document retrieval. Demonstrates ArcadeDB’s vector capabilities without requiring graph traversal — pure semantic search with filtering, faceting, and hybrid dense + sparse + keyword ranking.

Planned Features

Vector Similarity — Product and document embeddings with LSM_VECTOR, HNSW, and DiskANN indexes
Sparse Vector Retrieval — SPLADE / BM25 / BGE-M3 sparse embeddings via LSM_SPARSE_VECTOR (v26.5.1+)
Server-side Hybrid Fusion — vector.fuse combining dense + sparse + full-text in one query with RRF / DBSF / LINEAR strategies (v26.5.1+)
Full-Text Search — Lucene-backed keyword index, plugged in as a SEARCH_INDEX(…) source of vector.fuse
Document Model — Faceted filtering on product attributes
Python — Primary implementation language targeting data science and AI workflows

Reference Hybrid Query

Available since ArcadeDB v26.5.1.

SELECT expand(`vector.fuse`(
    `vector.neighbors`('Product[dense]', :queryVec, 50),
    `vector.sparseNeighbors`('Product[tokens,weights]', :qIdx, :qVal, 50),
    (SELECT @rid, $score FROM Product WHERE SEARCH_INDEX('Product[name]', :keywords) = true),
    { fusion: 'RRF', groupBy: 'category', groupSize: 1 }
)) LIMIT 20

This is the typical e-commerce shape: semantic similarity (dense embeddings) + exact-term matching (sparse) + keyword search (full-text), fused server-side, then diversified across product categories so the result page does not collapse to one category.

Target Scenarios

E-commerce product search with natural language queries
Document retrieval with semantic understanding
Hybrid search combining keyword matching and dense+sparse vector similarity
Multi-vector search (title embeddings + description embeddings + sparse term embeddings)