Vector index memory configuration

Vector indexes are based on Lucene. Lucene-backed vector index files are cached by the operating system’s filesystem cache rather than Neo4j page cache memory, as described in the Memory configuration section. For vector indexes, you must ensure that there is sufficient memory for Neo4j (JVM heap and Neo4j page cache) and that enough RAM remains available for the operating system’s filesystem cache. If insufficient RAM is left for that cache, the OS will read data from disk more often and vector search performance will degrade. Under broader memory pressure, the OS may also start swapping. Tools like iotop, or equivalent Linux I/O monitoring tools, can help you understand disk I/O usage.

Optimal memory configuration for Neo4j with vector indexes

To estimate a vector index’s filesystem cache requirements, you need to understand how Lucene stores the index. The main file types are:

.vex files for the HNSW graph
.vec files for the full-precision vector values
.veq files for scalar- or binary-quantized vector values

Lucene also creates metadata files, but these are small compared with the graph and vector data files. Depending on the segment size, Lucene can store these files as entries inside a compound .cfs file instead of as separate files on the filesystem. When inspecting an index that uses compound files, use a Lucene-aware tool that reports the entries inside each .cfs file, such as a tool based on Lucene’s CompoundDirectory API. A filesystem search for .vex, .vec, or .veq files alone undercounts the index.

For optimal performance, you want enough filesystem cache to hold the HNSW graph (.vex) and its associated vector values in memory. The associated vector values depend on the quantization used by the index: either the default binary or scalar (.veq) vectors, or full-precision unquantized (.vec) vectors. For example, for binary quantization, you want enough memory for the .vex and .veq files.

Quantization is useful when there is not enough memory to keep the full-precision vector values in the filesystem cache. Because quantization reduces search accuracy, especially with binary quantization, binary-quantized search results are by default rescored using the full-precision vector values from the .vec files. Only the vectors selected as candidates by the quantized search need to be read for rescoring. The .vec files, therefore, do not need to be cached, and reads for uncached values can be served from disk.

If the index has already been built, sum the sizes of the Lucene files or compound-file entries that need to be cached. Lucene indexes are made up of segments, and each segment has its own vector file set, so include the files for every segment:

For an unquantized index, sum the .vex and .vec files.
For a scalar-quantized index, sum the .vex and .veq files.
For a binary-quantized index, sum the .vex and .veq files.

Example 1. Measured vector index file sizes

The following table shows measured file sizes for indexes containing 1 million 768-dimensional FLOAT32 vectors. The indexes were built with Neo4j 2026.07.1 and vector.hnsw.m = 16. Lucene created between 313 and 327 segments per index, and the sizes include entries from every segment, including those stored in compound files.

Quantization HNSW graph Values used for search Full-precision values retained for rescoring Filesystem cache for the index

Quantization	HNSW graph	Values used for search	Full-precision values retained for rescoring	Filesystem cache for the index
None	46.7MB (`.vex`)	3.072GB (`.vec`)	Not applicable	3.119GB
Scalar	46.8MB (`.vex`)	784.0MB (`.veq`)	3.072GB (`.vec`)	830.8MB
Binary	46.7MB (`.vex`)	112.0MB (`.veq`)	3.072GB (`.vec`)	158.7MB

None

46.7MB (.vex)

3.072GB (.vec)

Not applicable

3.119GB

Scalar

46.8MB (.vex)

784.0MB (.veq)

3.072GB (.vec)

830.8MB

Binary

46.7MB (.vex)

112.0MB (.veq)

3.072GB (.vec)

158.7MB

The filesystem cache values include the HNSW graph and the values used for search. For quantized indexes, these filesystem cache values do not include the full-precision .vec files used to rescore candidates. The file sizes vary with Neo4j and Lucene versions, index settings, vector data, and segment layout, so measure the file sizes of an existing index rather than relying on these example values.

Estimating file sizes

When planning an index, you can estimate the sizes of these files as follows. The example uses the same 1 million 768-dimensional FLOAT32 vectors as in the previous measured vector index file sizes example.

Start by calculating the size of the full-precision vector values using the formula:
- full-precision vector values ≈ 4 bytes per dimension x dimension count x vector count
  full-precision vector values ≈ 4 bytes x 768 dimensions x 1,000,000 = 3.072 GB
  For Lucene-backed vector indexes, the dimension precision is 4 bytes even if the VECTOR type in Neo4j has a different precision.
If the index is using quantization, estimate the quantized vector data size using the formulas:
- scalar-quantized vector values ≈ full-precision vector values / 4
- binary-quantized vector values ≈ full-precision vector values / 32
  scalar-quantized vector values ≈ 3.072 GB / 4 = 768 MB binary-quantized vector values ≈ 3.072 GB / 32 = 96 MB
Estimate the size of the HNSW graph using the formula:
- HNSW graph size ≈ 8 bytes x vector count x HNSW_M
  HNSW graph size ≈ 8 bytes x 1,000,000 x 16 HNSW_M = 128 MB
  This formula is intended to provide an upper estimate. HNSW_M is the configured vector.hnsw.m value (default 16).
Calculate the filesystem cache requirement for the index by adding the vector-value and HNSW graph estimates:
- for a full-precision, unquantized index: 3.072 GB + 0.128 GB = 3.2 GB
- for a scalar-quantized index: 768 MB + 128 MB = 896 MB
- for a binary-quantized index with rescoring: 96 MB + 128 MB = 224 MB

Neo4j page cache depends on vector access pattern

Access pattern	Page cache guidance
Vectors are indexed and searched, but not returned or reused	Neo4j page cache can be sized mainly for the rest of the graph and any other accessed properties. The vector property values do not need to stay hot in Neo4j page cache if the query does not read them back from the graph store.
Vectors are returned, re-ranked, or reused after the index lookup	Neo4j page cache must also account for reading vector property values from the graph store. Size the page cache allocation accordingly.

Access pattern

Page cache guidance

Vectors are indexed and searched, but not returned or reused

Neo4j page cache can be sized mainly for the rest of the graph and any other accessed properties. The vector property values do not need to stay hot in Neo4j page cache if the query does not read them back from the graph store.

Vectors are returned, re-ranked, or reused after the index lookup

Neo4j page cache must also account for reading vector property values from the graph store. Size the page cache allocation accordingly.

In-index filtering

Additional properties declared on a vector index for in-index filtering are also stored in the Lucene index. During a filtered vector search, Lucene reads this data to determine which vectors are eligible for the search. For optimal filtered search performance, add the filter index data to the operating system’s filesystem cache estimate:

filesystem cache requirement ≈ vector index working set + filter index data

The Lucene files used for filtering depend on the property type. Numeric, temporal, and duration values use point indexes, which the current Lucene format stores in .kdd, .kdi, and .kdm files. String and boolean values use term dictionaries and postings, stored primarily in .tim, .tip, and .doc files. Some property types use both. As with the vector files, these can be entries inside compound .cfs files, and there is one set for each segment.

For capacity planning, estimate the filter index data by summing the following for every filterable property:

filter index data ≈ sum of (populated value count x (encoded value width + 8 bytes))

The populated value count is the number of indexed entities that have a value for the filterable property. The additional 8 bytes per populated value is a planning allowance for Lucene document associations and index structures, rather than a fixed storage cost. This provides a conservative estimate because Lucene compresses point values, document identifiers, repeated terms, and common string prefixes. Use the following encoded value widths:

Property type Encoded value width

Property type	Encoded value width
`BOOLEAN`	1 byte
`INTEGER` or `FLOAT`	8 bytes
`STRING`	Average UTF-8 encoded length of the values
`DATE`, `LOCAL DATETIME`, or `LOCAL TIME`	12 bytes
`ZONED TIME`	16 bytes
`ZONED DATETIME`	16 bytes plus the average UTF-8 encoded length of the zone ID
`DURATION`	32 bytes

BOOLEAN

1 byte

INTEGER or FLOAT

8 bytes

STRING

Average UTF-8 encoded length of the values

DATE, LOCAL DATETIME, or LOCAL TIME

12 bytes

ZONED TIME

16 bytes

ZONED DATETIME

16 bytes plus the average UTF-8 encoded length of the zone ID

DURATION

32 bytes

For example, an index expected to grow to 10 million vectors with a populated INTEGER filter and a populated STRING filter averaging 20 UTF-8 bytes per value requires the following filter index estimate:

10M x 8 + 8) + (20 + 8 = 440 MB

AuraDB instance sizing

For AuraDB-specific vector sizing and vector-optimized instance guidance, see Aura → Vector optimization.

Warming up the vector index

The vector index is loaded into the operating system’s filesystem cache as it is queried. The first queries may need to read more of the Lucene index from disk. As more of the index becomes resident in the filesystem cache, later queries avoid more disk reads and performance becomes more consistent.

Warm-up can therefore be done by issuing random queries against the index before serving production traffic. The number of queries required depends on the size of the index, the amount of RAM available, and how representative the warm-up queries are. As a starting point:

for a smaller index (up to 1M entries), around five random queries have worked well in testing
for a larger index, start with around 100 random queries and adjust based on observed disk I/O and latency on the target system

Use representative queries and monitor disk read activity during warm-up. Tools like iotop, or equivalent Linux I/O monitoring tools, can help show whether the operating system is still reading heavily from disk.

Glossary

allocator: A component in the cluster that allocates databases to servers according to the topology constraints specified and an allocation strategy.
asynchronous replication: Asynchronous replication is used by secondary copies to poll for new transactions, which means they cannot be guaranteed to have received the most recent transactions. This enables efficient scale-out of read-performance.
Aura instance: A fully-managed DBMS represented by a single instance ID, that is running in the Neo4j Aura cloud.
auto-commit transaction: An automatically committed transaction that contains a single query.
Bolt protocol: Bolt is a protocol used for interaction between Neo4j instances and drivers.
bookmark: A marker the client can request from the cluster to ensure that it is able to read its own writes so that the application’s state is consistent and only databases that have a copy of the bookmark are permitted to respond.
category (Bloom): A category is based on a node label and is defined in a Perspective as a way of visually distinguishing nodes with the same label(s).
causal consistency: All servers in a cluster agree on the order in which transactions take place. The position of a server on the causal chain can be guaranteed using a bookmark.
cluster: A Neo4j DBMS that spans multiple servers working together to increase fault tolerance and/or read scalability. Databases on a cluster may be configured to replicate across servers in the cluster thus achieving read scalability or high availability.
client application: Software that interacts with a Neo4j server.
commit: A commit is the successful completion of a transaction, which ensures durability of any changes made. For more details, visit Operations Manual → Transaction management.
composite database: Composite databases are the means to access partitioned graph data with a single Cypher query.
constraint: Constraints are sets of data modeling rules that ensure the data is consistent and reliable.
Cypher®: Neo4j’s graph query language.
data model: A data model defines how information is organized in a database. A good data model will make querying and understanding your data easier. In Neo4j, the data models have a graph structure.
database: A database is a container used by the DBMS to manage and store graph data. The physical structure of data is controlled by the database.
database vs graph: Databases are the physical containers of graph data. Graphs are the logical structure of data in Neo4j.
Database Management System: Database Management System, or DBMS, capable of managing multiple databases. A DBMS may run on a single server, or span several servers configured as a cluster.
database schema: The prescribed property existence and datatypes for nodes and relationships.
deallocate: An act of removing a database from a server or a server from a cluster without loss of data or reduced fault tolerance.
degree (of a node): The number of relationships of a specific node; loops are counted twice.
disaster recovery: A manual intervention to restore availability of a cluster, or databases within a cluster.
driver: A software library that provides access to Neo4j from a particular programming language.
election: In the event that the Raft leader becomes unresponsive, followers automatically trigger an election and vote for a new leader.
entity: A node or a relationship.
expression (Cypher): A component of a Cypher query which produces values. It may be used in projections, as a predicate, or when setting properties on graph elements.
fabric: Fabric is the architectural design of a unified system that provides a single access point to local or distributed graph data.
fault tolerance: A guarantee that a cluster can maintain a database’s persistence and availability in the event of one or more servers failing.
follower: A primary copy of a database acting as a follower, receives and acknowledges synchronous writes from the leader.
Generative AI (GenAI): A type of artificial intelligence (AI) system that generates text, images, or other media in response to prompts.
graph: A logical representation of a set of nodes where some pairs are connected by relationships.
index: Data structure that improves read performance of a database.
knowledge graph: A specific type of graph that has an organizing principle so that a user (or a computer system) can reason about the underlying data. The organizing principle provides an additional layer of structure that adds context to support knowledge discovery.
label: Marks a node as a member of a named and indexed subset. A node may be assigned zero or more labels.
leader: A single primary copy of a database is designated as the leader. It receives all write transactions from clients and replicates writes synchronously to followers and asynchronously to secondary copies of the database.
main database: In terms of Neo4j Enterprise Studio, the database(s) containing the user’s data. Can exist in the same Neo4j deployment as the tool asset database.
motif: A description of a specific pattern within a graph.
node: A node represents an entity or discrete object in your graph data model. Nodes can be connected by relationships, hold data in properties, and are classified by labels.
operator: A symbol representing a mathematical or logical operation.
parameter: Named value provided when running a Cypher statement.
path: A sequence of nodes and the relationships connecting them, that does not contain duplicate relationships. Several paths can match a pattern.
pattern: A specific arrangement of nodes and relationships that can be matched in a graph. A pattern follows a motif.
perspective (Bloom): A Perspective defines a certain business view or domain that can be found in the target Neo4j graph. A single Neo4j graph can be viewed through different Perspectives, each tailored for a different business purpose.
primary: A copy of the database that is able to process write transactions and is eligible to be elected as a leader. It participates in fault tolerant writes as it is part of the majority required to acknowledge and commit write transactions.
primary vs secondary: In a cluster, databases can operate in either primary or secondary mode. Primary databases are able to process write and read transactions, ensuring fault tolerance. Secondary databases are replicated asynchronously from primaries, and their main purpose is to provide read scaling within the cluster.
project (Aura): An isolated environment in the unified Aura console that contains its own database instances, configurations, and resources. Preceded by tenant in the classic Aura console.
property: Properties are key-value pairs that are used for storing data on nodes and relationships.
query (Cypher): A statement that retrieves or writes information to a database.
Raft group: A group of servers that are participating in hosting a particular database in primary mode.
Raft group member: A server that is participating in a Raft group. A server can be a member of one or more groups.
Raft log: A shared log between all Raft group members that is guaranteed to be consistently updated and viewed by those members. The log contains both database data and operational state of the Raft group.
Raft protocol: The networking mechanism that enables a database to replicate its data across multiple servers to give high availability for accessing the data and high durability to the data stored.
read scaling: Distributing query load by creating additional database copies hosted in secondary mode (read-only).
relationship: A relationship represents a connection between nodes in your graph data model. Relationships connect a source node to a target node, hold data in properties, and are classified by type.
secondary: An asynchronously replicated copy of the database that provides read scaling within the cluster.
seed: A seed is a database dump or a full backup used to create a database on a cluster. This is sometimes called seeding.
server: A physical machine, a virtual machine, or a container running an instance of Neo4j. Servers can be standalone or part of a cluster.
session: A causally linked sequence of transactions.
session consistency: An alternative name for Neo4j’s causal consistency.
standalone: A single server running Neo4j and not part of a cluster.
synchronous replication: Synchronous replication requires the leader primary to replicate a transaction and block the commit until a quorum of the follower primaries acknowledges that the transaction is successfully replicated. Once the transaction is replicated, the commit is allowed to proceed. This ensures data durability and consistency within the cluster.
system database: A database used by Neo4j to store system information.
tenant (Aura): An isolated environment in the classic Aura console that contains its own database instances, configurations, and resources. Replaced by project in the unified Aura console.
tool asset database: In terms of Neo4j Enterprise Studio, the database where tools' assets are stored. This can be in the same Neo4j deployment as the main database(s) or in a separate deployment.
topology: A configuration that describes how the copies of a database should be spread across the servers in a cluster, see primary mode and secondary mode.
transaction: A transaction comprises a unit of work performed against a database. It is treated in a coherent and reliable way, independent of other transactions. Transactions comply with the ACID consistency model (atomic, consistent, isolated, and durable).