2025

A Complete
Guide to
Resource
Optimization
Resource optimization isn’t just a choice - it’s a necessity.

Mastering a few tricks can make all the difference.

Table of Contents
03 What’s in This Guide? 13 Optimize Queries by Properly Filtering

05 When Should You Optimize? 14 Ultimate Precision With Hybrid Search

06 Optimization is a Balancing Act 15 Reranking for Relevance

07 Configure Indexing for Faster Searches 16 Reduce Overhead With Batch Processing

08 Vector Index Optimization Parameters 17 Storage Management

09 Scalar Quantization 19 Plan Your Resource Capacity: RAM vs Disk

10 Binary Quantization 20 Optimization Cheatsheet

11 Multitenancy 21 Conclusion

12 Custom Sharding
What’s in This Guide? Insider tips to help you fine-tune your vector database for peak performance.

Is your AI application
resource-hungry and hard
to scale?
You're not alone. Many organizations struggle with unoptimized
implementations, leading to wasted resources and limited real-world impact.
But here's the good news: by fine-tuning your setup, you can transform your
vector database into a powerhouse of efficiency and performance.

Whether you're running a small-scale experiment or a production-grade

application, these insights will help you refine your database for maximum
efficiency.

A Complete Guide to Resource Optimization 03

What’s in This Guide? Insider tips to help you fine-tune your vector database for peak performance.

Highlights Include:

Why does Resource

Optimization Matter? Resource Management Strategies:
Trying to scale your GenAI app on a budget? We will show you how to avoid
wasting compute resources and get the maximum return on your investment.
Every query counts when scaling AI-driven applications. Optimizing your
vector database isn’t just about saving costs—it’s about building systems that
deliver top-notch performance while being lean and adaptable. By optimizing
resources, you’ll ensure your applications remain competitive, scalable, and
Performance Improvement Tricks:
ready for real-world use.

We’ll dive into advanced techniques like indexing, compression, and

partitioning. Our tips will help you get better results at scale, while reducing
Qdrant powers millions of GenAI apps, not just due to its high speeds, but also
total resource expenditure.
because of its strong focus on performance optimization. As a vector
database and a search engine, it is designed to be highly configurable even in
the most extreme business cases. In this guide you will learn how our
performance optimization features solve different challenges and when to use
them. Query Optimization Methods:
Improving your vector database setup isn’t just about saving costs. We’ll
Let’s break down the strategies that will help you fine-tune your vector show you how to build search systems that deliver consistently high precision
while staying adaptable.
database.

A Complete Guide to Resource Optimization 04

When Should You Optimize? The critical moments that drive your need for database efficiency.

Optimization can significantly enhance your system’s speed and responsiveness.

When Should You Optimize?

01 02
When You Scale Up If You’re Facing Budget Constraints
As data grows and the requests surge, optimizing Strike the perfect balance between performance and
resource usage ensures your systems stay responsive cost, cutting unnecessary expenses while maintaining
and cost-efficient, even under heavy loads. essential capabilities.

03 04
You Need Better Performance When System Stability is Paramount
If you’re noticing slow query speeds, latency issues, or To manage high-traffic environments you will need to
frequent timeouts, it’s time to fine-tune your resource prevent crashes or failures caused by resource
allocation. exhaustion.

A Complete Guide to Resource Optimization 05

Optimization is a Balancing Act Understanding the tradeoffs between resources is the key to mastery.

Optimization is a High Speed

Balancing Act
First, choose the Optimization Strategy that best fits your
Intended Result to properly balance your resource expenditure.

Learn more about Qdrant’s Optimization Methods

Intended Result Optimization Strategy Pick Two

High Precision + Low Memory On-Disk Indexing

Low Memory + High Speed Quantization

Low Memory High Precision
High Precision + High Speed RAM Storage + Quantization
Figure 1: Different use cases require different balances between memory usage,
search speed, and precision.
Latency vs Throughput Segment Configuration

A Complete Guide to Resource Optimization 06

Configure Indexing for Faster Searches Customize key parameters to balance speed, accuracy, and resource use.

Configure Indexing entry

layer 2 nearest
point on

for Faster Searches

neighbour on
layer l = 2
layer l =2

What is it?
layer 1 nearest

neighbour on

A vector index is the central location where Qdrant calculates layer l = 1 entry

point on

vector similarity. It is the backbone of your search process, layer l = 1

retrieving relevant results from vast amounts of data.

Qdrant uses the HNSW (Hierarchical Navigable Small World layer 0

Graph) algorithm as its dense vector index, which is both query

powerful and scalable.

nearest

neighbour on

entry

layer l = 0
point on

layer l = 0
Figure 2: A sample HNSW vector index with three layers. Follow the blue arrow
on the top layer to see how a query travels throughout the database index. The
closest result is on the bottom level, nearest to the gray query point. Figure 2

A Complete Guide to Resource Optimization 07

Vector Index Optimization Parameters

Vector Index m (Edges per Node)

Optimization
This controls the number of edges in the graph. A higher value
enhances search accuracy but demands more memory and
build time. Fine-tune this to balance memory usage and

Parameters
precision.

These parameters give you the flexibility to fine-tune Qdrant’s

ef_construct (Index Build Range)
performance for your specific workload. You can modify them
directly in Qdrant's configuration files or at the collection and This parameter sets how many neighbors are considered during
index construction. A larger value improves the accuracy of the
named vector levels for more granular control.
index but increases the build time. Use this to customize your
indexing speed versus quality.
Learn more about Indexing

ef (Search Range)
This determines how many neighbors are evaluated during a
search query. You can adjust this to balance query speed and
accuracy.

A Complete Guide to Resource Optimization 08

Scalar Quantization Reducing the number of bits used to represent each vector in the database.

Scalar Quantization
Scalar quantization strikes an excellent balance between
compression and performance, making it the go-to choice for float 32 vector [ range -1 to +1 ]

most use cases.

-0.9 0.2 0.8 -0.5 -1.0 0.6 0.3 -0.8 0.9 -0.1 40 bytes
Learn more about Scalar Quantization

Memory usage will drop:

4x

Compression cuts memory usage by a factor of 4. Qdrant reduction

compresses 32-bit floating-point values (float32) into 8-bit
unsigned integers (uint8).

-115 25 102 -64 -128 77 38 -102 115 -12 10 bytes

Accuracy loss is minimal
Converting from float32 to uint8 introduces a small loss in int8 vector [ range -128 to +127 ]
precision. Typical error rates remain below 1%, making this
method highly efficient.
Figure 3: The top example shows a float32 vector with a size of 40 bytes. Converting it to int8 format
Best for specific use cases: reduces its size by a factor of four, while preserving the original representation of the user data.
To be used with high-dimensional vectors where minor
accuracy losses are acceptable.

A Complete Guide to Resource Optimization 09

Binary Quantization Reducing each vector in the database down to a single bit.

Binary Quantization
Original Vector
Binary quantization is ideal for high-dimensional datasets and
compatible embedding models, where compression and speed
0.98 -0.08 0.72 -0.43 -0.12 0.01 size= 6144 bytes
are paramount.

Learn more about Binary Quantization

...1536 dimensions in total... 32x reduction

Efficient similarity calculations:
Emulates Hamming distance through dot product
comparisons, making it fast and effective.

1 0 1 0 0 1 size= 192 bytes

Perfect for high-dimensional vectors:
Works well with embedding models like OpenAI’s text- compressed representation

embedding-ada-002 or Cohere’s embed-english-v2.0.

Precision management: Figure 4: This method causes maximum compression. It reduces memory usage by 32x and speeds up
Consider rescoring or oversampling to offset precision loss. searches by up to 40x.

A Complete Guide to Resource Optimization 10

M ultitenancy A single collection can handle the complexity of global multi-user systems.

Multitenancy
A single collection can handle the complexity of global
multi-user systems.
collection_name=”tenant_data”
Learn more about Multitenancy

Why Multitenancy?

Logical Isolation:
Tenant 1 Tenant 2
Ensures each tenant’s data remains separate while
residing in the same collection.

id=1 ,
 id=2 ,
id=3 ,

Minimized Overhead: payload={“group_id”: “tenant_1”} ,
payload={“group_id”: “tenant_1”} ,
payload={“group_id”: “tenant_2”} ,

Reduces resource consumption compared to vector=[0.9, 0.1, 0.1] , vector=[0.1, 0.9, 0.1] , vector=[0.1, 0.1, 0.9] ,

maintaining separate collections for each user.

Scalability:
H andles high user volumes without compromising Figure 5: Each individual vector is assigned a specific payload that denotes which tenant it belongs to. This is how

performance. a large number of different tenants can share a single Qdrant collection

A Complete Guide to Resource Optimization 11

Custom Sharding Control where your data is placed inside of a single Qdrant collection.

Custom Update

Sharding
shard 1: canada shard 3: usa shard 2: germany shard 4: india

Retrieve

Figure 6

User-Defined Sharding is particularly useful in multi-tenant setups, as it enables the isolation of

each tenant’s data within separate shards, ensuring better organization and enhanced data
security.

Figure 6: Users can both upsert and query shards that are relevant to them, all within the same collection. Regional
sharding can help avoid cross-continental traffic.

Sharding is a critical strategy in Qdrant for splitting collections into smaller units, called shards, to
efficiently distribute data across multiple nodes. It’s a powerful tool for improving scalability and
maintaining performance in large-scale system

Learn more about Sharding

A Complete Guide to Resource Optimization 12

Optimize Queries by Properly Filtering Enhancing search performance, speed and accuracy.

When dealing with large datasets, it’s impractical—and inefficient—to search through every

Optimize Queries data point. Instead, you can significantly improve performance by applying filters on specific
payload fields. Filtering allows you to narrow down the search space by excluding irrelevant data

by Properly points, thereby reducing the computational load and focusing only on the most relevant subset
of your dataset.

Filtering Default Vector Index

Query Vector
Introducing Filters

Filtered out vector

Filterable Vector Index

Additional Link

01 02 03
Entry point Entry point Entry point

Figure 7: This technique builds additional links (orange) between leftover data points. The filtered points which stay
behind are now traversable once again. Qdrant uses special category-based methods to connect these data points.

Qdrant’s filterable vector index works fast and it is the best method of capturing all available
results.

Learn more about Filtering

A Complete Guide to Resource Optimization 13

Ultimate Precision With Hybrid Search

Ultimate Precision With

Hybrid Search Dense Results

Sparse Results
Hybrid search combines keyword filtering with vector similarity search,
enabling faster and more precise results. Keywords help narrow down the
dataset quickly, while vector similarity ensures semantic accuracy. Hybrid Normalization
search in Qdrant combines results from two types of data:
Dense Results
Dense vector search:
Finds results based on semantic similarity using vector embeddings. Sparse Results

Sparse vector search:

Matches specific terms or keywords using traditional inverted indices, Fusion
similar to full-text search.
Mixture
Hybrid search in Qdrant fuses results from different methods, like dense
vectors and keyword search, by normalizing their scores and combining
them. This creates a final score that balances both semantic relevance Figure 8: To combine Sparse and Dense vectors, Qdrant uses Reciprocal Rank Fusion as a
and exact matches, improving overall search accuracy. method of normalising results.

Learn more about Hybrid Search

A Complete Guide to Resource Optimization 14

Reranking for Relevance Weighing hybrid search results by importance.

Reranking for Relevance

D1 D2 D3
When both sparse and dense vectors are used in a query, Qdrant retrieves D2
candidate results based on their combined relevance. However, because the D4 ... DN
initial ranking is based on approximate similarity, reranking can be applied to D2
refine the result set and improve accuracy.
D1 D2 D3 DN
After you receive the initial hybrid search results, you can rerank them using
Reranking
late interaction embeddings for maximum precision.
D4 ... DN D4

Considerations
...
Reranking can be computationally expensive, so aim for a balance between D1 D2 D3
relevance and speed. The first step is to retrieve the relevant documents and D1
then use reranking. D4 ... DN

Regularly evaluate your reranking models to avoid overfitting and make

timely adjustments to maintain performance.
Figure 9: Reranking adjusts the order of search results based on additional criteria, ensuring
Learn more about Reranking the most relevant results are prioritized.

A Complete Guide to Resource Optimization 15

Reduce Overhead With Batch Processing Lower transaction requirements and speed up data processing.

Reduce Overhead With

Batch Processing Data 1

Batch processing consolidates multiple operations into a single execution NOT

Qdrant
Qdrant

Data 2 Shards
cycle. It’s an effective strategy for both data insertion and query execution. BATCHED Cluster Peers

Batch Update Data 3

Instead of inserting vectors individually, group them into larger batches to

minimize the number of database transactions and the overhead of frequent
writes. This reduces write operations and ensures faster data ingestion.

Learn more about Batch Update Data 1

Batch Queries BATCHED Data 2

Qdrant
Qdrant

Shards
Cluster Peers
Similarly, you can batch multiple queries together rather than executing
them one by one. This reduces the number of round trips to the database,
Data 3
optimizing performance and reducing latency. Batch queries are particularly
useful when processing a large number of similar queries or when handling
multiple user requests simultaneously.

Learn more about Batch Queries Figure 10: Batch Processing Example

A Complete Guide to Resource Optimization 16

Storage Management Choose between in-memory and on-disk storage to meet your performance and scalability goals.

Storage
Management How it works
All data is stored in RAM, providing the fastest access times for queries and
operations.
As your data scales, effective resource management
becomes crucial to keeping costs low while ensuring
your application remains reliable and performant.

When to use it
This setup is ideal for applications where performance is critical, and your RAM
Qdrant supports two main methods for storing vectors capacity can accommodate all vectors and their payloads.
and payloads: InMemory and OnDisk/Memmap.

Advantages
Learn more about Storage
You can reach the maximum speed for vector/payload queries and updates.

Limitations
RAM usage can become a bottleneck as your dataset grows.

InMemory
OnDisk / Memmap

A Complete Guide to Resource Optimization 17

Storage Management Choose between in-memory and on-disk storage to meet your performance and scalability goals.

Storage How it works

Management Instead of loading all data into memory, memmap storage maps data files directly
to a virtual address space on disk. The system's page cache handles data access,
making it highly efficient.

As your data scales, effective resource management

becomes crucial to keeping costs low while ensuring When to use it
your application remains reliable and performant.

Perfect for storing large collections that exceed your available RAM while still
maintaining near in-memory performance when enough RAM is available.
Qdrant supports two main methods for storing vectors
and payloads: InMemory and OnDisk/Memmap.
Advantages
Learn more about Storage Balances performance and memory usage, allowing you to work with datasets
larger than your physical RAM. For larger datasets or scenarios where memory is
limited, OnDisk storage is more suitable. This method significantly reduces
memory usage by storing data on disk.

Limitations
InMemory
Slightly slower than pure in-memory storage but significantly more scalable.

OnDisk / Memmap

A Complete Guide to Resource Optimization 18

Plan Your Resource Capacity: RAM vs Disk Strike a balance between RAM, disk, and hardware type for efficient scaling.

Plan Your Resource

When scaling a Q drant cluster, selecting the right hardware is critical to
achieving a balance between RAM and disk storage, tailored to the specific
needs of your dataset.

Capacity: RAM vs Disk Learn more about Capacity Planning

More RAM or a Bigger Disk? Which Disk Type Should I Use?

RAM: SSD:
Crucial for fast access to frequently used data, such as indexed vectors. The Recommended for optimal performance, particularly for workloads involving
amount of RAM required can be estimated based on your dataset size and random reads and writes. SSDs can significantly enhance query response
dimensionality. For example, storing 1 million vectors with 1024 dimensions times when the data is stored on disk.
would require approximately 5.72 GB of RAM.

Disk: HDD:
Suitable for less frequently accessed data, such as payloads and non-critical While more cost-effective, HDDs are slower and can negatively impact
information. Disk-backed storage reduces memory demands but can performance, especially for large datasets or applications requiring high-
introduce slight latency. speed access.

A Complete Guide to Resource Optimization 19

Optimization Cheatsheet Address common database bottlenecks and enhance scalability with proven fixes.

Optimization Cheatsheet
Connecting the techniques discussed earlier to common challenges in managing vector
databases can provide practical strategies to optimize performance and scalability.

Reduce Excessive Memory Manage Large Datasets in FixSlow Performance and Query Avoid High Costs by Reducing
Consumption Distributed Systems Timeouts Data Overlap
Large datasets, particularly vector data, can lead As the volume of vector data increases, storing and Performance bottlenecks during querying and retrieval In multi-user environments, isolating user-specific data is
to significant memory usage, straining resources managing it across distributed nodes becomes can result in slow responses or timeouts, particularly in crucial to ensure data security and prevent interference
and degrading system performance. challenging, often leading to slower data retrieval times systems with poor optimization or inefficient storage between users. However, managing this isolation across
and scaling issues. methods. multiple nodes can lead to increased operational costs.
Offloading Load from RAM to SSDs:
Extend memory capacity by storing data on Partition and distribute the dataset across multiple Optimize & refine query structures: Implement a multitenant architecture:
SSDs, which offer a cost-effective way to reduce nodes in a cluster to enhance scalability and improve Leverage approximate nearest neighbor (ANN) This allows user data to be efficiently isolated while
RAM reliance while maintaining reasonable retrieval performance. algorithms to reduce query latency and enhance sharing a common infrastructure. This approach
access speeds. response times. minimizes redundancy by storing user data logically
Load Balancing in requests:
separated within the same system, reducing both
Vector Quantization Techniques: Distribute incoming requests evenly across nodes to Cache frequently accessed data in memory:
resource usage and costs.
Compress vectors to lower memory ensure consistent performance and avoid overloading This minimizes the load on the database and speeds up
requirements without compromising accuracy any single node. query execution. Build access control mechanisms:
significantly. Techniques like scalar or product This ensures that user data remains isolated, even within
quantization can be particularly effective. a shared infrastructure.

Indexing Parameter Optimization:

Adjust parameters such as the number of
clusters or the indexing strategy to balance
memory usage and system performance
efficiently.

A Complete Guide to Resource Optimization 20

Conclusion We hope this guide helps you optimize your resources and fine-tune your vector database
for peak performance. What’s next? You can read about these topics with the links
throughout the guide, or if you are ready to start optimizing:

Deploy locally Try Managed Cloud Try Hybrid Cloud

Get started with our Quick Start Guide. Start prototyping then scale with your needs once production- Bring your own cluster to access robust cloud infrastructure and
ready. scaling capabilities.
Access GitHub Repo
Try free Get started

Working with 10,000,000+ vectors and want to make sure you are set up for success? Talk to sales
About Qdrant:
Qdrant is the leading, high-performance, scalable, open-source vector database and search

engine, essential for building the next generation of AI/ML applications.

Qdrant is able to handle billions of vectors, supports the matching of semantically complex

objects, and is implemented in Rust for performance, memory safety, and scale. Recently,

the company was recognized among the top 10 startups on Sifted’s 2024 B2B SaaS Rising

100, which annually ranks Europe's most promising B2B SaaS startups valued under $1bn.

https://qdrant.tech/

https://discord.com/invite/qdrant

https://github.com/qdrant/qdrant