Elastic + Qdrant Hybrid Search Demo

A demo application showcasing hybrid search for e-commerce by combining Elasticsearch (keyword/BM25 search) with Qdrant (vector/semantic search). Uses the H&M product dataset with pre-computed dense embeddings.

Architecture

• Frontend: Next.js with React and Tailwind CSS
• Backend: FastAPI (Python)
• Search: Elasticsearch 9.0 + Qdrant
• Embeddings: BAAI/bge-small-en-v1.5 (384 dimensions)

Quick Start

1. Start the services

docker compose --profile debug up --build

This starts:

• Elasticsearch (port 9200)
• Kibana (port 5601) - for debugging/visualization
• Qdrant (port 6333)
• Backend API (port 8000)
• Frontend (port 3000)

Note: Omit --profile debug if you don't need Kibana.

2. Load data into Elasticsearch

Once the services are healthy, load the product data:

cd scripts
pip install elasticsearch polars
python load_elastic.py

The script will download the H&M product dataset from Hugging Face and index it into Elasticsearch.

3. Migrate vectors to Qdrant (optional)

To enable joint search mode (ES + Qdrant with RRF fusion):

export ES_HOST=http://localhost:9200
export ES_USERNAME=elastic
export ES_PASSWORD=changeme

./scripts/migrate_dense_vectors_to_qdrant.sh

4. Use the app

Open http://localhost:3000 and search for products.

Search Modes

• Elasticsearch: Hybrid search combining BM25 keyword matching with dense vector similarity
• Joint: Queries both Elasticsearch and Qdrant, merges results using Reciprocal Rank Fusion (RRF)

Environment Variables

Variable	Default	Description
ES_PASSWORD	changeme	Elasticsearch password
ES_VERSION	9.0.0	Elasticsearch/Kibana version
ES_API_KEY	-	Optional API key for ES auth
QDRANT_API_KEY	-	Optional API key for Qdrant