Indexers

Indexers

VortexDB supports multiple indexing algorithms, each optimized for different use cases. The index determines how vectors are organized for similarity search.

Flat Index

The Flat index performs brute-force exhaustive search by computing distances to every vector.

When to Use

• Dataset size: < 10,000 vectors
• Requirements: Need exact/guaranteed results
• Use cases: Testing, prototyping, small production workloads

INDEX_TYPE=flat

KD-Tree Index

The KD-Tree (k-dimensional tree) is a space-partitioning data structure that recursively divides the vector space.

At each level, the tree splits data along a different dimension (cycling through x, y, z, …).

When to Use

• Vector dimensions: < 20 dimensions
• Dataset size: Thousands to hundreds of thousands of vectors
• Use cases: Geographic data, low-dimensional embeddings, spatial queries

INDEX_TYPE=kdtree

Caution

For high-dimensional embeddings (e.g., 384, 768, 1536 dimensions common in ML), KD-Tree is not recommended. Use HNSW instead.

HNSW Index

HNSW (Hierarchical Navigable Small World) is a state-of-the-art approximate nearest neighbor algorithm based on proximity graphs. It constructs a multi-layered graph where each layer represents a different level of granularity, enabling efficient navigation from coarse to fine-grained similarity search.

Check out this blog post for more theoretical details and this blog covering implementation in VortexDB.

When to Use

• Vector dimensions: Any, but especially > 20 dimensions
• Dataset size: 100,000+ vectors
• Use cases: Semantic search, recommendation systems, RAG applications

INDEX_TYPE=hnsw

Distance Metrics

All indexes support four distance/similarity metrics:

Cosine

Measures the angle between two vectors, ignoring magnitude.

$d = 1 - \frac{A \cdot B}{\|A\| \|B\|}$

• Range: 0 (identical) to 2 (opposite)
• Best for: Text embeddings, normalized vectors

similarity=Similarity.COSINE

Euclidean

L2 distance—the straight-line distance between points.

$d = \sqrt{\sum_{i=1}^{n} (a_i - b_i)^2}$

• Range: 0 (identical) to ∞
• Best for: General purpose

similarity=Similarity.EUCLIDEAN

Manhattan

L1 distance—the sum of absolute differences.

$d = \sum_{i=1}^{n} |a_i - b_i|$

• Range: 0 (identical) to ∞
• Best for: Sparse vectors, grid-based data

similarity=Similarity.MANHATTAN

Hamming

Counts positions where elements differ.

• Range: 0 (identical) to n (completely different)
• Best for: Binary vectors, categorical data

similarity=Similarity.HAMMING

Choosing an Index

Use this decision tree:

Vectors	Dimensions	Recommended Index
< 10,000	Any	Flat
10K - 100K	< 20	KD-Tree
10K - 100K	≥ 20	HNSW
> 100K	Any	HNSW

Configuration Example

# For a semantic search application with OpenAI embeddings
DIMENSION=1536
INDEX_TYPE=hnsw
STORAGE_TYPE=rocksdb

# For a small prototype with sentence-transformers
DIMENSION=384
INDEX_TYPE=flat
STORAGE_TYPE=inmemory

Next Steps

Snapshots

Learn how to backup and restore your index

API Reference

Explore the complete API