Vector Engine

The vector engine powers semantic search — nearest-neighbor retrieval over high-dimensional embeddings. It uses a custom HNSW index with multiple quantization levels and hardware-accelerated distance math.

When to Use

• Semantic search over embeddings (text, images, audio)
• RAG pipelines for AI agents
• Recommendation systems
• Similarity matching and deduplication

Key Features

• HNSW index — Multi-layer proximity graph. Construction at full precision (FP32/FP16) for structural integrity; traversal on quantized payloads for cache residency.
• Quantization — SQ8 (~4x memory reduction), PQ (~4-8x), IVF-PQ (~16 bytes/vector for 100M+ datasets).
• Adaptive pre-filtering — Roaring Bitmap-based filtering. Automatic strategy selection: pre-filter (selective), post-filter (broad), or brute-force.
• Distance metrics — L2, cosine, inner product, Manhattan, Chebyshev, Hamming, Jaccard, Pearson.
• Cross-engine fusion — Combine with graph (GraphRAG), full-text (hybrid BM25+vector), or spatial filtering.

SQL Usage

-- Create a collection with a vector index
CREATE COLLECTION articles;
CREATE VECTOR INDEX idx_embed ON articles METRIC cosine DIM 384;

-- Insert with embedding
INSERT INTO articles { title: 'Understanding Transformers', embedding: [0.12, -0.34, 0.56, ...] };

-- Nearest neighbor search
SEARCH articles USING VECTOR(embedding, ARRAY[0.1, 0.3, -0.2, ...], 10);

-- Filtered vector search
SELECT title, vector_distance(embedding, ARRAY[0.1, 0.3, ...]) AS score
FROM articles
WHERE category = 'machine-learning'
  AND id IN (SEARCH articles USING VECTOR(embedding, ARRAY[0.1, 0.3, ...], 10));

-- Hybrid BM25 + vector (RRF fusion)
SELECT title, rrf_score(
    vector_distance(embedding, $query_vec),
    bm25_score(body, 'transformer attention')
) AS score
FROM articles
LIMIT 10;

Quantization & Storage Precision

Storage Precision (storage_dtype — raw HNSW structural storage):

Quantization (quantization — index-level compression, independent of storage_dtype):

Codes are traversed for speed, then top candidates are re-ranked against full-precision vectors. Memory is the per-vector code size at 384 dimensions; smaller codes trade recall for footprint and scale.

How It Works

Vectors are indexed in the HNSW graph at full precision. During search, quantized copies are traversed for speed, then top candidates are re-ranked against full-precision vectors. When metadata filters are present, the engine builds a Roaring Bitmap of matching IDs and selects the optimal strategy based on selectivity.