ANN Backends

GrafitoDB supports multiple Approximate Nearest Neighbor (ANN) backends for vector search.

Available Backends

Backend	Install	Best For	Persistence
FAISS	pip install grafitodb[faiss]	Production, most features	✅ Yes
Annoy	pip install grafitodb[annoy]	Large datasets, memory-mapped	✅ Yes
LEANN	pip install grafitodb[leann]	Fast builds, small datasets	✅ Yes
HNSWlib	pip install grafitodb[hnswlib]	High recall	⚠️ Soft delete
USearch	pip install grafitodb[usearch]	Modern, fast	✅ Yes
Voyager	pip install grafitodb[voyager]	Spotify's library	✅ Yes
Brute Force	Built-in	Small datasets, exact search	❌ No

FAISS

Facebook AI Similarity Search. Most feature-complete backend.

Installation

pip install grafito[faiss]
# Or with conda:
conda install -c pytorch faiss-cpu

Methods

# Flat (exact search)
db.create_vector_index(
    name='docs',
    dim=384,
    backend='faiss',
    method='flat',
    options={'metric': 'l2'}
)

# IVF (inverted file index)
db.create_vector_index(
    name='docs',
    dim=384,
    backend='faiss',
    method='ivf_flat',
    options={'nlist': 100, 'nprobe': 10}
)

# HNSW (hierarchical navigable small world)
db.create_vector_index(
    name='docs',
    dim=384,
    backend='faiss',
    method='hnsw',
    options={'hnsw_m': 16, 'ef_construction': 200}
)

Persistence

db.create_vector_index(
    name='docs',
    dim=384,
    backend='faiss',
    method='flat',
    options={'index_path': 'indexes/docs.faiss'}
)

Annoy

Spotify's Approximate Nearest Neighbors Oh Yeah.

Installation

pip install grafito[annoy]

Characteristics

• Memory-mapped indexes (shareable between processes)
• Good for read-heavy workloads
• Static indexes (rebuild to add vectors)

db.create_vector_index(
    name='docs',
    dim=384,
    backend='annoy',
    method='annoy',
    options={
        'metric': 'angular',  # or 'euclidean', 'manhattan'
        'n_trees': 50,        # More trees = more accurate, slower build
        'index_path': 'indexes/docs.annoy'
    }
)

LEANN

Lightweight Efficient Approximate Nearest Neighbors.

Installation

pip install grafito[leann]

Characteristics

• Fast index building
• Good for small to medium datasets
• Auto-build control

db.create_vector_index(
    name='docs',
    dim=384,
    backend='leann',
    method='leann',
    options={
        'metric': 'l2',
        'auto_build': False,  # Disable auto-build
        'index_path': 'indexes/docs.leann'
    }
)

# Add embeddings...

# Manual rebuild
db.rebuild_vector_index('docs')

HNSWlib

Hierarchical Navigable Small World implementation.

Installation

pip install grafito[hnswlib]

Characteristics

• High recall rates
• Supports incremental adds
• Soft deletes (need rebuild to purge)

db.create_vector_index(
    name='docs',
    dim=384,
    backend='hnswlib',
    method='hnswlib',
    options={
        'metric': 'l2',
        'M': 16,
        'ef_construction': 200,
        'ef': 50
    }
)

Note

HNSWlib uses soft deletes. Call rebuild_vector_index() to fully remove deleted vectors.

USearch

Modern FAISS alternative by Unum.

Installation

pip install grafito[usearch]

Characteristics

• Faster than FAISS for many workloads
• Smaller memory footprint
• Native persistence

db.create_vector_index(
    name='docs',
    dim=384,
    backend='usearch',
    method='usearch',
    options={
        'metric': 'cos',
        'connectivity': 16,
        'expansion_add': 128,
        'expansion_search': 64
    }
)

Voyager

Spotify's latest ANN library.

Installation

pip install grafito[voyager]

Characteristics

• Multi-threaded index building
• Good for large-scale search
• EFIL (Enhanced Forward Index Layout)

db.create_vector_index(
    name='docs',
    dim=384,
    backend='voyager',
    method='voyager',
    options={
        'space': 'cosine',  # or 'l2', 'ip'
        'M': 16,
        'ef_construction': 200,
        'index_path': 'indexes/docs.voyager'
    }
)

Brute Force

Built-in exact search for small datasets.

# No extra installation needed
db.create_vector_index(
    name='docs',
    dim=384,
    backend='bruteforce',
    method='bruteforce',
    options={'metric': 'l2'}
)

Use for: - Datasets < 1000 vectors - When exact results are required - Testing and debugging

Backend Comparison

Feature	FAISS	Annoy	LEANN	HNSWlib	USearch	Voyager
Incremental adds	✅	❌	✅	✅	✅	✅
GPU support	✅*	❌	❌	❌	❌	❌
Memory mapped	✅	✅	❌	❌	✅	✅
Cosine similarity	✅	✅	✅	✅	✅	✅
Inner product	✅	❌	✅	✅	✅	✅
L2 distance	✅	✅	✅	✅	✅	✅

*Requires faiss-gpu (conda)

Choosing a Backend

Small Dataset (< 10K vectors)

# Brute force or FAISS flat
db.create_vector_index(..., backend='bruteforce', method='bruteforce')
# or
db.create_vector_index(..., backend='faiss', method='flat')

Medium Dataset (10K - 100K)

# FAISS IVF or HNSW
db.create_vector_index(..., backend='faiss', method='ivf_flat')
# or
db.create_vector_index(..., backend='faiss', method='hnsw')

Large Dataset (> 100K)

# Annoy (memory-mapped) or Voyager
db.create_vector_index(..., backend='annoy', method='annoy')
# or
db.create_vector_index(..., backend='voyager', method='voyager')

Read-Heavy Workloads

# Annoy with memory mapping
db.create_vector_index(..., backend='annoy', method='annoy',
                       options={'index_path': 'shared.annoy'})

Write-Heavy Workloads

# FAISS HNSW or HNSWlib
db.create_vector_index(..., backend='faiss', method='hnsw')

Checking Available Backends

from grafito.integrations import available_vector_backends

print(available_vector_backends())
# ['bruteforce', 'faiss', 'annoy', 'leann', 'hnswlib', 'usearch', 'voyager']