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

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
Euclidean
Manhattan
Hamming

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

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

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

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

Getting Started

Concepts

Indexers

Indexers

Flat Index

When to Use

KD-Tree Index

When to Use

HNSW Index

When to Use

Distance Metrics

Choosing an Index

Configuration Example

Next Steps

Snapshots

API Reference

​Indexers

​Flat Index

​When to Use

​KD-Tree Index

​When to Use

​HNSW Index

​When to Use

​Distance Metrics

​Choosing an Index

​Configuration Example

​Next Steps

Snapshots

API Reference

Indexers

Flat Index

When to Use

KD-Tree Index

When to Use

HNSW Index

When to Use

Distance Metrics

Choosing an Index

Configuration Example

Next Steps