Memory optimization

Learn how to optimize memory consumption in Redis vector sets

Vector set is a new data type that is currently in preview and may be subject to change.

Overview

Redis vector sets are efficient, but vector similarity indexing and graph traversal require memory tradeoffs. This guide helps you manage memory use through quantization, graph tuning, and attribute choices.

Quantization modes

Vector sets support three quantization levels:

Mode	Memory usage	Recall	Notes
`Q8`	4x smaller	High	Default, fast and accurate
`BIN`	32x smaller	Lower	Fastest, best for coarse search
`NOQUANT`	Full size	Highest	Best precision, slowest

Use Q8 unless your use case demands either ultra-precision (use NOQUANT) or ultra-efficiency (use BIN).

Graph structure memory

HNSW graphs store multiple connections per node. Each node:

Has an average of M * 2 + M * 0.33 pointers (default M = 16).
Stores pointers using 8 bytes each.
Allocates ~1.33 layers per node.

A single node with M = 64 may consume ~1 KB in links alone.

To reduce memory:

Lower M to shrink per-node connections.
Avoid unnecessarily large values for M unless recall needs to be improved.

Attribute and label size

Each node stores:

A string label (element name)
Optional JSON attribute string

Tips:

Use short, fixed-length strings for labels.
Keep attribute JSON minimal and flat. For example, use {"year":2020} instead of nested data.

Vector dimension

High-dimensional vectors increase storage:

300 components at FP32 = 1200 bytes/vector
300 components at Q8 = 300 bytes/vector

You can reduce this using the REDUCE option during VADD, which applies random projection:

>_ Redis CLI

>VADD setNotReduced VALUES 300 ... element
(integer) 1
> VDIM setNotReduced
(integer) 300

>VADD setReduced REDUCE 100 VALUES 300 ... element
(integer) 1
> VDIM setReduced
(integer) 100

Python

"""
Code samples for Vector set doc pages:
    https://redis.io/docs/latest/develop/data-types/vector-sets/
"""

import redis

from redis.commands.vectorset.commands import (
    QuantizationOptions
)

r = redis.Redis(decode_responses=True)


res1 = r.vset().vadd("points", [1.0, 1.0], "pt:A")
print(res1)  # >>> 1

res2 = r.vset().vadd("points", [-1.0, -1.0], "pt:B")
print(res2)  # >>> 1

res3 = r.vset().vadd("points", [-1.0, 1.0], "pt:C")
print(res3)  # >>> 1

res4 = r.vset().vadd("points", [1.0, -1.0], "pt:D")
print(res4)  # >>> 1

res5 = r.vset().vadd("points", [1.0, 0], "pt:E")
print(res5)  # >>> 1

res6 = r.type("points")
print(res6)  # >>> vectorset

res7 = r.vset().vcard("points")
print(res7)  # >>> 5

res8 = r.vset().vdim("points")
print(res8)  # >>> 2

res9 = r.vset().vemb("points", "pt:A")
print(res9)  # >>> [0.9999999403953552, 0.9999999403953552]

res10 = r.vset().vemb("points", "pt:B")
print(res10)  # >>> [-0.9999999403953552, -0.9999999403953552]

res11 = r.vset().vemb("points", "pt:C")
print(res11)  # >>> [-0.9999999403953552, 0.9999999403953552]

res12 = r.vset().vemb("points", "pt:D")
print(res12)  # >>> [0.9999999403953552, -0.9999999403953552]

res13 = r.vset().vemb("points", "pt:E")
print(res13)  # >>> [1, 0]

res14 = r.vset().vsetattr("points", "pt:A", {
    "name": "Point A",
    "description": "First point added"
})
print(res14)  # >>> 1

res15 = r.vset().vgetattr("points", "pt:A")
print(res15)
# >>> {'name': 'Point A', 'description': 'First point added'}

res16 = r.vset().vsetattr("points", "pt:A", "")
print(res16)  # >>> 1

res17 = r.vset().vgetattr("points", "pt:A")
print(res17)  # >>> None

res18 = r.vset().vadd("points", [0, 0], "pt:F")
print(res18)  # >>> 1

res19 = r.vset().vcard("points")
print(res19)  # >>> 6

res20 = r.vset().vrem("points", "pt:F")
print(res20)  # >>> 1

res21 = r.vset().vcard("points")
print(res21)  # >>> 5

res22 = r.vset().vsim("points", [0.9, 0.1])
print(res22)
# >>> ['pt:E', 'pt:A', 'pt:D', 'pt:C', 'pt:B']

res23 = r.vset().vsim(
    "points", "pt:A",
    with_scores=True,
    count=4
)
print(res23)
# >>> {'pt:A': 1.0, 'pt:E': 0.8535534143447876, 'pt:D': 0.5, 'pt:C': 0.5}

res24 = r.vset().vsetattr("points", "pt:A", {
    "size": "large",
    "price": 18.99
})
print(res24)  # >>> 1

res25 = r.vset().vsetattr("points", "pt:B", {
    "size": "large",
    "price": 35.99
})
print(res25)  # >>> 1

res26 = r.vset().vsetattr("points", "pt:C", {
    "size": "large",
    "price": 25.99
})
print(res26)  # >>> 1

res27 = r.vset().vsetattr("points", "pt:D", {
    "size": "small",
    "price": 21.00
})
print(res27)  # >>> 1

res28 = r.vset().vsetattr("points", "pt:E", {
    "size": "small",
    "price": 17.75
})
print(res28)  # >>> 1

# Return elements in order of distance from point A whose
# `size` attribute is `large`.
res29 = r.vset().vsim(
    "points", "pt:A",
    filter='.size == "large"'
)
print(res29)  # >>> ['pt:A', 'pt:C', 'pt:B']

# Return elements in order of distance from point A whose size is
# `large` and whose price is greater than 20.00.
res30 = r.vset().vsim(
    "points", "pt:A",
    filter='.size == "large" && .price > 20.00'
)
print(res30)  # >>> ['pt:C', 'pt:B']

# Import `QuantizationOptions` enum using:
#
# from redis.commands.vectorset.commands import (
#   QuantizationOptions
# )
res31 = r.vset().vadd(
    "quantSetQ8", [1.262185, 1.958231],
    "quantElement",
    quantization=QuantizationOptions.Q8
)
print(res31)  # >>> 1

res32 = r.vset().vemb("quantSetQ8", "quantElement")
print(f"Q8: {res32}")
# >>> Q8: [1.2643694877624512, 1.958230972290039]

res33 = r.vset().vadd(
    "quantSetNoQ", [1.262185, 1.958231],
    "quantElement",
    quantization=QuantizationOptions.NOQUANT
)
print(res33)  # >>> 1

res34 = r.vset().vemb("quantSetNoQ", "quantElement")
print(f"NOQUANT: {res34}")
# >>> NOQUANT: [1.262184977531433, 1.958230972290039]

res35 = r.vset().vadd(
    "quantSetBin", [1.262185, 1.958231],
    "quantElement",
    quantization=QuantizationOptions.BIN
)
print(res35)  # >>> 1

res36 = r.vset().vemb("quantSetBin", "quantElement")
print(f"BIN: {res36}")
# >>> BIN: [1, 1]

# Create a list of 300 arbitrary values.
values = [x / 299 for x in range(300)]

res37 = r.vset().vadd(
    "setNotReduced",
    values,
    "element"
)
print(res37)  # >>> 1

res38 = r.vset().vdim("setNotReduced")
print(res38)  # >>> 300

res39 = r.vset().vadd(
    "setReduced",
    values,
    "element",
    reduce_dim=100
)
print(res39)  # >>> 1

res40 = r.vset().vdim("setReduced")  # >>> 100
print(res40)

This projects a 300-dimensional vector into 100 dimensions, reducing size and improving speed at the cost of some recall.

Summary

Strategy	Effect
Use `Q8`	Best tradeoff for most use cases
Use `BIN`	Minimal memory, fastest search
Lower `M`	Shrinks HNSW link graph size
Reduce dimensions	Cuts memory per vector
Minimize JSON	Smaller attributes, less memory per node

Products

Tools

Get Redis

Connect

Learn

Latest

See how it works