Vector Search

Vector search enables similarity-based queries on high-dimensional numerical data, supporting use cases such as semantic search, recommendation systems, and image retrieval. Cassandra 5.0 introduced native vector support with the VECTOR data type and approximate nearest neighbor (ANN) queries via Storage-Attached Indexes (SAI).

---

Overview

What is Vector Search?

Vector search finds data items similar to a query vector based on mathematical distance or similarity measures. Unlike traditional equality or range queries that match exact values, vector search returns results ordered by proximity in high-dimensional space.

Traditional Query:  WHERE status = 'active'     → Exact match
Vector Query:       ORDER BY embedding ANN OF ? → Similarity ranked

How Vectors Represent Data

Machine learning models transform unstructured data (text, images, audio) into fixed-dimension numerical vectors called embeddings. Similar items produce vectors that are mathematically close to each other.

Data Type	Embedding Model Examples	Typical Dimensions
Text	OpenAI ada-002, Cohere, Sentence-BERT	384 - 3072
Images	ResNet, CLIP, ViT	512 - 2048
Audio	Wav2Vec, OpenL3	512 - 1024
Multimodal	CLIP, BLIP	512 - 1024

Embedding Generation

Cassandra stores and queries vectors but does not generate embeddings. Applications must use external ML models or embedding APIs to convert source data into vectors before insertion.

---

Behavioral Guarantees

What Vector Search Guarantees

• Vector columns store fixed-dimension floating-point arrays
• SAI vector indexes support approximate nearest neighbor (ANN) queries
• ORDER BY ... ANN OF returns results ordered by similarity to the query vector
• LIMIT controls the maximum number of results returned
• Similarity functions (similarity_cosine, similarity_euclidean, similarity_dot_product) compute distances consistently

What Vector Search Does NOT Guarantee

Undefined Behavior

The following behaviors are undefined and must not be relied upon:

• Exact nearest neighbors: ANN queries return approximate results; the true k-nearest neighbors may not be included
• Result stability: Identical queries may return slightly different results due to ANN approximation
• Cross-partition ordering: When querying multiple partitions, global ordering is approximate
• Recall guarantees: The percentage of true nearest neighbors found varies by index configuration and data distribution
• Performance with filters: Combining ANN with WHERE clauses may significantly impact performance
• Dimension mismatch handling: Query vectors with wrong dimensions cause errors; no automatic padding or truncation

ANN Query Contract

Aspect	Guarantee
Result ordering	Approximate similarity order (best matches first)
Result count	At most LIMIT rows returned
Index requirement	SAI index with 'similarity_function' option required
Query vector dimension	Must match column dimension exactly

Similarity Function Contract

Function	Range	Best Match	Use Case
similarity_cosine	-1 to 1	1 (identical direction)	Normalized embeddings, text similarity
similarity_euclidean	0 to ∞	0 (identical)	Spatial data, unnormalized vectors
similarity_dot_product	-∞ to ∞	Higher is more similar	Pre-normalized vectors, performance

Index Build Contract

State	Query Behavior
Building	ANN queries may fail or return incomplete results
Built	ANN queries return approximate results from all indexed data
Missing	ANN queries fail with error

Failure Semantics

Failure Mode	Outcome	Client Action
Dimension mismatch	Query rejected	Ensure query vector matches column dimension
Missing ANN index	Query fails	Create SAI index with similarity function
Index building	Partial or no results	Wait for index build completion
Query timeout	Partial results possible	Reduce LIMIT or add partition constraints

Version-Specific Behavior

Version	Behavior
5.0+	VECTOR type and ANN queries introduced (CEP-30)
5.0+	SAI vector indexes with similarity functions

---

Vector Data Type

Syntax

VECTOR<element_type, dimension>

Parameter	Description	Constraints
element_type	Numeric type for elements	FLOAT (recommended), DOUBLE
dimension	Number of elements	Fixed at table creation, typically 128-3072

Creating Tables with Vectors

-- Semantic search table
CREATE TABLE documents (
    doc_id UUID,
    title TEXT,
    content TEXT,
    embedding VECTOR<FLOAT, 1536>,
    created_at TIMESTAMP,
    PRIMARY KEY (doc_id)
);

-- Product recommendations
CREATE TABLE products (
    product_id UUID,
    category TEXT,
    name TEXT,
    description TEXT,
    image_embedding VECTOR<FLOAT, 512>,
    text_embedding VECTOR<FLOAT, 768>,
    PRIMARY KEY ((category), product_id)
);

-- Multi-modal search
CREATE TABLE media_assets (
    asset_id UUID,
    asset_type TEXT,
    visual_embedding VECTOR<FLOAT, 1024>,
    audio_embedding VECTOR<FLOAT, 512>,
    PRIMARY KEY (asset_id)
);

Dimension Immutability

Vector dimensions cannot be changed after table creation. Migrating to a different embedding model with different dimensions requires creating a new table and re-embedding all data.

---

Inserting Vector Data

Basic Insert

-- Insert with vector literal
INSERT INTO documents (doc_id, title, content, embedding, created_at)
VALUES (
    uuid(),
    'Introduction to Cassandra',
    'Apache Cassandra is a distributed database...',
    [0.023, -0.142, 0.089, ...],  -- 1536 float values
    toTimestamp(now())
);

Parameterized Insert (Recommended)

-- Prepared statement with vector parameter
INSERT INTO documents (doc_id, title, content, embedding, created_at)
VALUES (?, ?, ?, ?, ?);

Driver examples:

# Python driver
from cassandra.cluster import Cluster

cluster = Cluster(['127.0.0.1'])
session = cluster.connect('my_keyspace')

# Embedding from ML model (list of floats)
embedding = model.encode("Apache Cassandra is a distributed database...")

prepared = session.prepare("""
    INSERT INTO documents (doc_id, title, content, embedding, created_at)
    VALUES (?, ?, ?, ?, toTimestamp(now()))
""")

session.execute(prepared, [
    uuid.uuid4(),
    'Introduction to Cassandra',
    'Apache Cassandra is a distributed database...',
    embedding  # Pass as list or array
])

// Java driver
import com.datastax.oss.driver.api.core.data.CqlVector;

PreparedStatement prepared = session.prepare(
    "INSERT INTO documents (doc_id, title, content, embedding, created_at) " +
    "VALUES (?, ?, ?, ?, toTimestamp(now()))"
);

float[] embeddingArray = model.encode("Apache Cassandra is a distributed database...");
CqlVector<Float> embedding = CqlVector.newInstance(embeddingArray);

session.execute(prepared.bind(
    UUID.randomUUID(),
    "Introduction to Cassandra",
    "Apache Cassandra is a distributed database...",
    embedding
));

Batch Embedding

For bulk data loading, batch embedding requests to the ML model for efficiency. Most embedding APIs support batching multiple texts in a single request.

Updating Vectors

-- Update existing vector
UPDATE documents
SET embedding = [0.031, -0.128, 0.095, ...]
WHERE doc_id = ?;

Update Implications

Updating vectors triggers index updates. For frequently changing vectors, consider the write amplification impact on SAI indexes.

---

Vector Indexing

Vector similarity search requires an SAI index with a specified similarity function.

Creating Vector Indexes

-- Cosine similarity (default, recommended for text embeddings)
CREATE CUSTOM INDEX documents_embedding_idx ON documents (embedding)
    USING 'StorageAttachedIndex';

-- Explicit similarity function
CREATE CUSTOM INDEX ON documents (embedding)
    USING 'StorageAttachedIndex'
    WITH OPTIONS = {'similarity_function': 'cosine'};

-- Euclidean distance (L2)
CREATE CUSTOM INDEX ON products (image_embedding)
    USING 'StorageAttachedIndex'
    WITH OPTIONS = {'similarity_function': 'euclidean'};

-- Dot product
CREATE CUSTOM INDEX ON products (text_embedding)
    USING 'StorageAttachedIndex'
    WITH OPTIONS = {'similarity_function': 'dot_product'};

Similarity Functions

Function	Formula	Range	Best For
cosine	1 - cos(θ)	0 to 2	Normalized embeddings, text similarity
euclidean	√Σ(aᵢ - bᵢ)²	0 to ∞	Spatial data, image features
dot_product	Σ(aᵢ × bᵢ)	-∞ to ∞	Pre-normalized vectors, performance

Choosing Similarity Functions

• Cosine: Default choice for most text embeddings. Measures angle between vectors, insensitive to magnitude.
• Euclidean: Use when absolute distances matter, such as geographic or spatial embeddings.
• Dot product: Fastest computation. Requires normalized vectors (magnitude = 1) for meaningful results.

Similarity Function Consistency

The similarity function must match how the embedding model was trained. Most text embedding models (OpenAI, Cohere, Sentence-BERT) are optimized for cosine similarity.

---

Querying Vectors

Approximate Nearest Neighbor (ANN) Search

The ORDER BY ... ANN OF clause performs similarity search:

-- Find 10 most similar documents
SELECT doc_id, title, similarity_cosine(embedding, ?) AS similarity
FROM documents
ORDER BY embedding ANN OF ?
LIMIT 10;

LIMIT Required

ANN queries require a LIMIT clause. Queries without LIMIT will fail. The maximum allowed LIMIT is 1000.

Query Execution Flow

Similarity Score Functions

Retrieve similarity scores in the result set:

-- Cosine similarity score
SELECT doc_id, title, similarity_cosine(embedding, ?) AS score
FROM documents
ORDER BY embedding ANN OF ?
LIMIT 10;

-- Euclidean distance score
SELECT product_id, name, similarity_euclidean(image_embedding, ?) AS distance
FROM products
ORDER BY image_embedding ANN OF ?
LIMIT 10;

-- Dot product score
SELECT doc_id, similarity_dot_product(embedding, ?) AS score
FROM documents
ORDER BY embedding ANN OF ?
LIMIT 10;

Function	Return Value	Interpretation
similarity_cosine	0 to 1	Higher = more similar
similarity_euclidean	0 to ∞	Lower = more similar
similarity_dot_product	-∞ to ∞	Higher = more similar

Filtered Vector Search

Combine vector search with scalar predicates:

-- Vector search with equality filter
SELECT doc_id, title, similarity_cosine(embedding, ?) AS score
FROM documents
WHERE category = 'technical'
ORDER BY embedding ANN OF ?
LIMIT 10;

-- Vector search with partition key (most efficient)
SELECT product_id, name, similarity_cosine(image_embedding, ?) AS score
FROM products
WHERE category = 'electronics'
ORDER BY image_embedding ANN OF ?
LIMIT 20;

-- Multiple filters
SELECT doc_id, title
FROM documents
WHERE category = 'science'
  AND created_at > '2024-01-01'
ORDER BY embedding ANN OF ?
LIMIT 10;

Filter Order Matters

For optimal performance, design partition keys to support common filter combinations. Queries that include the partition key with vector search perform significantly better than global vector scans.

Filter Selectivity

SAI applies filters before or after vector search based on estimated selectivity. Highly selective filters (matching few rows) are applied first, reducing the vector search space.

---

Complete Application Example

Semantic Document Search

-- Schema
CREATE TABLE IF NOT EXISTS knowledge_base (
    doc_id UUID,
    title TEXT,
    content TEXT,
    category TEXT,
    embedding VECTOR<FLOAT, 1536>,
    created_at TIMESTAMP,
    PRIMARY KEY (doc_id)
);

-- Indexes
CREATE CUSTOM INDEX ON knowledge_base (embedding)
    USING 'StorageAttachedIndex'
    WITH OPTIONS = {'similarity_function': 'cosine'};

CREATE CUSTOM INDEX ON knowledge_base (category)
    USING 'StorageAttachedIndex';

CREATE CUSTOM INDEX ON knowledge_base (created_at)
    USING 'StorageAttachedIndex';

# Python application
from cassandra.cluster import Cluster
import openai

class SemanticSearch:
    def __init__(self):
        self.cluster = Cluster(['127.0.0.1'])
        self.session = self.cluster.connect('my_keyspace')
        self.openai = openai.Client()

        # Prepare statements
        self.insert_stmt = self.session.prepare("""
            INSERT INTO knowledge_base
            (doc_id, title, content, category, embedding, created_at)
            VALUES (?, ?, ?, ?, ?, toTimestamp(now()))
        """)

        self.search_stmt = self.session.prepare("""
            SELECT doc_id, title, content, similarity_cosine(embedding, ?) AS score
            FROM knowledge_base
            ORDER BY embedding ANN OF ?
            LIMIT ?
        """)

        self.filtered_search_stmt = self.session.prepare("""
            SELECT doc_id, title, content, similarity_cosine(embedding, ?) AS score
            FROM knowledge_base
            WHERE category = ?
            ORDER BY embedding ANN OF ?
            LIMIT ?
        """)

    def get_embedding(self, text):
        """Generate embedding using OpenAI API"""
        response = self.openai.embeddings.create(
            model="text-embedding-ada-002",
            input=text
        )
        return response.data[0].embedding

    def insert_document(self, doc_id, title, content, category):
        """Insert document with generated embedding"""
        embedding = self.get_embedding(f"{title} {content}")
        self.session.execute(self.insert_stmt, [
            doc_id, title, content, category, embedding
        ])

    def search(self, query, limit=10, category=None):
        """Semantic search with optional category filter"""
        query_embedding = self.get_embedding(query)

        if category:
            rows = self.session.execute(
                self.filtered_search_stmt,
                [query_embedding, category, query_embedding, limit]
            )
        else:
            rows = self.session.execute(
                self.search_stmt,
                [query_embedding, query_embedding, limit]
            )

        return [
            {
                'doc_id': row.doc_id,
                'title': row.title,
                'content': row.content,
                'score': row.score
            }
            for row in rows
        ]

# Usage
search = SemanticSearch()

# Insert documents
search.insert_document(
    uuid.uuid4(),
    "Cassandra Data Modeling",
    "Learn how to design tables for query patterns...",
    "database"
)

# Search
results = search.search("How do I design Cassandra tables?", limit=5)
for result in results:
    print(f"{result['title']} (score: {result['score']:.4f})")

---

Strengths

Native Database Integration

Vector search is integrated directly into Cassandra's storage and query engine:

• Single system for both vector and scalar data
• Transactional consistency with other column updates
• No separate vector database infrastructure required
• Familiar CQL interface

Distributed Scalability

Inherits Cassandra's distributed architecture:

• Vectors partitioned and replicated like other data
• Horizontal scaling by adding nodes
• No single point of failure
• Multi-datacenter replication supported

Combined Queries

Filter vector results using standard CQL predicates:

• Combine ANN search with equality, range, and CONTAINS filters
• Leverage existing SAI indexes on scalar columns
• Single query for hybrid search patterns

Storage Efficiency

SAI vector indexes are storage-attached:

• Indexes compact with SSTables
• No separate index storage management
• Consistent with Cassandra's LSM-tree architecture

---

Limitations

Approximate Results

ANN search returns approximate nearest neighbors, not exact results:

• Results are ranked by estimated similarity
• True nearest neighbors may be missed (recall < 100%)
• Recall accuracy depends on index parameters and data distribution
• No guarantee of deterministic ordering for equal similarity scores

Dimension Constraints

Vector dimensions have practical limits:

• Maximum recommended dimension: 4096
• Higher dimensions increase storage and computation costs
• Dimension reduction may be needed for very large embeddings
• Dimension is fixed at table creation

Query Restrictions

Vector queries have specific requirements:

• LIMIT clause is mandatory (max 1000)
• ORDER BY ... ANN OF required for similarity search
• Cannot combine with GROUP BY or aggregations
• Cannot use in materialized view definitions

Memory Requirements

Vector indexes consume significant memory:

• HNSW graph structures reside partially in memory
• Memory usage scales with vector count and dimensions
• Large vector tables may require memory tuning
• Monitor heap usage during index builds

Index Build Time

Building vector indexes on existing data takes time:

• Proportional to vector count and dimensions
• CPU-intensive graph construction
• Monitor with nodetool compactionstats
• Consider off-peak index creation for large tables

No Cross-Partition Vector Aggregation

Vector search within partitions differs from global search:

• Global ANN scans all partitions (resource-intensive)
• Partition-restricted ANN is more efficient
• Design partition keys to align with common query patterns

---

Performance Tuning

Index Configuration

-- Tune graph construction parameters (advanced)
CREATE CUSTOM INDEX ON documents (embedding)
    USING 'StorageAttachedIndex'
    WITH OPTIONS = {
        'similarity_function': 'cosine'
        -- Additional SAI tuning options may be available
        -- in future Cassandra versions
    };

Query Optimization

Technique	Impact	Implementation
Include partition key	High	Design schema for partition-restricted vector queries
Reduce LIMIT	Medium	Request only needed results
Pre-filter with selective predicates	Medium	Add SAI indexes on filter columns
Normalize vectors	Low	Pre-normalize for dot_product similarity

Capacity Planning

Factor	Estimation
Vector storage	rows × dimensions × 4 bytes (FLOAT)
Index overhead	~1.5x vector storage
Memory per node	Monitor during index build, tune heap
Query latency	Benchmark with production-like data

Benchmarking

Always benchmark vector operations with production-representative data volumes and query patterns. Performance characteristics vary significantly based on:

• Vector dimensions
• Dataset size
• Query concurrency
• Cluster configuration

---

Use Cases

Semantic Search

Search documents by meaning rather than keywords:

-- User searches for "database performance issues"
-- Returns documents about optimization, tuning, slow queries, etc.
SELECT doc_id, title, similarity_cosine(embedding, ?) AS relevance
FROM documentation
ORDER BY embedding ANN OF ?
LIMIT 20;

Recommendation Systems

Find similar products or content:

-- Given a product the user viewed, find similar products
SELECT product_id, name, similarity_cosine(embedding,
    (SELECT embedding FROM products WHERE product_id = ?)) AS similarity
FROM products
WHERE category = ?
ORDER BY embedding ANN OF (SELECT embedding FROM products WHERE product_id = ?)
LIMIT 10;

Subquery Limitation

The above subquery syntax is illustrative. In practice, retrieve the source embedding first, then use it as a parameter.

Image Similarity

Find visually similar images:

-- Reverse image search
SELECT image_id, filename, similarity_euclidean(visual_embedding, ?) AS distance
FROM image_catalog
ORDER BY visual_embedding ANN OF ?
LIMIT 25;

Anomaly Detection

Identify outliers by finding items distant from cluster centroids:

-- Find transactions dissimilar to normal patterns
SELECT transaction_id, amount,
       similarity_cosine(feature_vector, ?) AS normality_score
FROM transactions
WHERE account_id = ?
ORDER BY feature_vector ANN OF ?
LIMIT 100;
-- Filter for low similarity scores in application

RAG (Retrieval-Augmented Generation)

Retrieve context for LLM prompts:

-- Find relevant context chunks for RAG
SELECT chunk_id, content, similarity_cosine(embedding, ?) AS relevance
FROM knowledge_chunks
WHERE domain = 'product_documentation'
ORDER BY embedding ANN OF ?
LIMIT 5;
-- Pass retrieved content to LLM as context

---

Migration and Compatibility

Version Requirements

Feature	Minimum Version
VECTOR data type	Cassandra 5.0
SAI vector index	Cassandra 5.0
similarity_* functions	Cassandra 5.0
ORDER BY ... ANN OF	Cassandra 5.0

Pre-5.0 Alternatives

Cassandra versions before 5.0 do not support native vector search. Alternatives include:

• External vector databases (Pinecone, Weaviate, Milvus)
• DataStax Astra DB (managed Cassandra with vector support)
• Application-level vector search with stored BLOBs

Schema Evolution

When changing embedding models:

-- 1. Add new vector column
ALTER TABLE documents ADD embedding_v2 VECTOR<FLOAT, 768>;

-- 2. Create new index
CREATE CUSTOM INDEX ON documents (embedding_v2)
    USING 'StorageAttachedIndex';

-- 3. Backfill embeddings (application-side)
-- 4. Update queries to use new column
-- 5. Drop old column and index when migration complete
DROP INDEX documents_embedding_idx;
ALTER TABLE documents DROP embedding;

---

Related Documentation

• Vector Data Type - DDL syntax for vector columns
• SAI Architecture - Index internals
• Secondary Index Queries - General index query patterns
• Data Modeling - Schema design principles