Hybrid Semantic Vector + Keyword Search RAM Sizing (Arroy & HNSW 768d/1536d)

Vector Embedding Dimensionality & Binary Quantization Math (MeiliSizer Guide)

Uncompressed 1536-dimension float32 vectors (text-embedding-3-small) consume 6.14 KB per document before graph edges — over 7.2 GB per 1M items! Switching to 768-dimension embeddings (BGE-base / multilingual-e5) or enabling Meilisearch v1.11+ Arroy binary quantization compresses vector storage to ~0.95–1.2 GB per 1M documents while retaining 96%+ semantic recall when fused with keyword BM25 (`semanticRatio: 0.5`).

Meilisearch v1.11+ Hybrid Embedder + Binary Quantization Configuration

// PATCH /indexes/products/settings — Hybrid Keyword + Binary Quantized Vectors
{
  "embedders": {
    "catalog_semantic": {
      "source": "openAi",
      "model": "text-embedding-3-small",
      "dimensions": 768,
      "binaryQuantized": true,
      "documentTemplate": "{{doc.title}} - {{doc.category}} - {{doc.summary}}"
    }
  }
}
// Query with semanticRatio: 0.45 (55% exact keyword + 45% vector concept)

MeiliSizer RAM & NVMe Tier Matrix for HYBRID-VECTOR-HNSW

MeiliSizer Production Checklist: HYBRID-VECTOR-HNSW

• Truncate Matryoshka embeddings (like text-embedding-3-small) from 1536d to 768d or 512d to cut vector RAM in half with <1% NDCG loss.
• Enable binaryQuantized: true on catalogs exceeding 1M items — keyword BM25 reranking restores exact precision on top hits.
• Keep documentTemplate concise (< 250 tokens) to slash embedding generation latency and API token bills.
• Provision dedicated physical VDS cores (UltaHost VDS or Vultr HF) with AVX2/AVX-512 SIMD instructions for fast cosine/Hamming distance scans.

Load Workload in MeiliSizer Matrix →