Embedding Generation

Overview

Generate embeddings from text, images, and multimodal data with intelligent caching.

Text Embeddings

Generate embeddings from text using the embed_text function. NeuronDB supports multiple embedding models and automatically caches results for performance.

Basic Text Embedding

-- Basic text embedding (uses default model)
SELECT 
    'Hello, world!' AS text,
    embed_text('Hello, world!') AS embedding,
    vector_dims(embed_text('Hello, world!')) AS dimensions;

-- Verify embedding is not null
SELECT 
    embed_text('Hello, world!') IS NOT NULL AS not_null,
    vector_dims(embed_text('Hello, world!')) AS dims;

Function Signature:

embed_text( text TEXT,                    -- Input text model_name TEXT DEFAULT NULL  -- Optional model name ) RETURNS VECTOR                  -- Returns embedding vector

Specify Embedding Model

-- Use specific embedding model
SELECT 
    embed_text('Test text', 'sentence-transformers/all-MiniLM-L6-v2') AS embedding,
    vector_dims(embed_text('Test text', 'sentence-transformers/all-MiniLM-L6-v2')) AS dims;

-- Common models:
-- 'all-MiniLM-L6-v2' (default, 384 dimensions)
-- 'sentence-transformers/all-MiniLM-L6-v2' (384 dimensions)
-- 'all-mpnet-base-v2' (768 dimensions)
-- 'text-embedding-ada-002' (1536 dimensions, requires API key)

Embedding Consistency

Same text produces the same embedding (cached):

-- Same text produces same embedding
WITH embeddings AS (
    SELECT
        embed_text('Consistency test') AS vec1,
        embed_text('Consistency test') AS vec2
)
SELECT
    vector_dims(vec1) = vector_dims(vec2) AS dims_match,
    vec1 <-> vec2 AS distance  -- Should be 0 or very close
FROM embeddings;

Unicode and Special Characters

-- Unicode and special characters are supported
SELECT 
    vector_dims(embed_text('Hello 世界 🌍')) AS unicode_dims,
    vector_dims(embed_text('Text with "quotes" and ''apostrophes''')) AS special_chars_dims;

Long Text

-- Long text is automatically handled
SELECT 
    vector_dims(embed_text(repeat('This is a long text. ', 100))) AS long_text_dims;

Batch Generation

Generate embeddings for multiple texts efficiently using batch functions. Batch processing is faster and more efficient than individual calls.

-- Batch embedding generation
-- Returns: array of vectors
SELECT embed_text_batch(
    ARRAY['First text', 'Second text', 'Third text']
) AS embeddings;

-- Use with table data
WITH texts AS (
    SELECT ARRAY_AGG(content) AS text_array
    FROM documents
    LIMIT 100
)
SELECT 
    embed_text_batch(text_array) AS batch_embeddings
FROM texts;

-- Process embeddings from batch
SELECT 
    unnest(embed_text_batch(ARRAY['text1', 'text2', 'text3'])) AS embedding;

Function Signature:

embed_text_batch( texts TEXT[],              -- Array of text strings model_name TEXT DEFAULT NULL ) RETURNS VECTOR[]             -- Returns array of vectors

Performance: Batch generation is significantly faster than individual embed_text() calls, especially when GPU acceleration is enabled.

Caching

Embeddings are automatically cached to improve performance. The cache stores embeddings by text content and model name, avoiding redundant computation.

-- View embedding cache statistics
SELECT * FROM neurondb.embedding_cache_stats;

-- View cached embeddings
SELECT 
    cache_key,
    model_name,
    vector_dims(embedding) AS dims,
    created_at
FROM neurondb.embedding_cache
ORDER BY created_at DESC
LIMIT 10;

-- Check cache hit rate
SELECT 
    cache_hits,
    cache_misses,
    ROUND(100.0 * cache_hits / NULLIF(cache_hits + cache_misses, 0), 2) AS hit_rate_percent
FROM neurondb.embedding_cache_stats;

Cache Configuration:

• Cache is enabled by default
• Cache key is based on text content and model name
• Cache size is configurable via neurondb.embedding_cache_size
• Cache automatically evicts least recently used entries when full

Configuration

Configure embedding generation via PostgreSQL GUC settings:

-- Set default embedding model
SET neurondb.default_embedding_model = 'all-MiniLM-L6-v2';

-- Configure LLM API key for external models (Hugging Face)
SET neurondb.llm_api_key = 'your-api-key';

-- Enable GPU acceleration for embeddings
SET neurondb.gpu_enabled = true;

-- Configure cache size (number of entries)
SET neurondb.embedding_cache_size = 10000;

Note: If LLM is not configured, embed_text() gracefully returns zero vectors as fallback. Configure API keys or enable GPU for real embeddings.

Learn More

For detailed documentation on embedding models, providers, caching strategies, and multimodal embeddings, visit: Embedding Generation Documentation

Related Topics

• Model Management - Manage embedding models
• Vector Search - Index and search embeddings