shreyask/qmd-web

QMD Web Demo

In-browser hybrid search pipeline using WebGPU + Transformers.js v4.

Demonstrates the full QMD search pipeline running entirely in your browser:

1. 1.Query Expansion — Qwen3 1.7B generates HyDE, semantic, and keyword variants
2. 2.Parallel Search — BM25 keyword search + vector similarity search
3. 3.Reciprocal Rank Fusion — Merges results from multiple search backends
4. 4.LLM Reranking — Qwen3 Reranker 0.6B scores document relevance
5. 5.Score Blending — Position-aware combination of RRF and reranker scores

Requirements

• —Chrome 113+ or Edge 113+ (WebGPU required)
• —~2.5GB model download on first visit (cached for subsequent visits)

Models

• —embeddinggemma-300M — Embeddings
• —Qwen3-Reranker-0.6B — Reranking
• —qmd-query-expansion-1.7B — Query expansion

Based on QMD by Tobi Lütke.