Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/spiceai/spiceai/llms.txt

Use this file to discover all available pages before exploring further.

Spice provides enterprise-grade search capabilities combining vector similarity, full-text, and keyword search for both structured and unstructured data.

​
Search Types

Spice supports three primary search methods:
  1. Vector Similarity Search - Semantic search using embeddings and distance metrics
  2. Full-Text Search - BM25-powered text search with Tantivy
  3. Keyword Search - Traditional exact and partial keyword matching
Combine multiple search methods using Reciprocal Rank Fusion (RRF) to achieve better relevance than any single method alone. Hybrid search merges results from vector and text search, reranking by relative position rather than raw scores. All search capabilities are exposed through SQL using User-Defined Table Functions (UDTFs):

​
Vector Search UDTF

SELECT * FROM vector_search(
  table_name,
  'search query',
  limit => 10
);

​
Text Search UDTF

SELECT * FROM text_search(
  table_name,
  'search query',
  limit => 10
);
These UDTFs integrate seamlessly with standard SQL operations: 
SELECT id, title, _score
FROM vector_search(documents, 'machine learning basics')
WHERE publish_date > '2024-01-01'
ORDER BY _score DESC
LIMIT 5;

​
Vector Storage Options

Spice supports multiple vector storage backends:
  • Amazon S3 Vectors - Petabyte-scale vector storage (recommended for production)
  • pgvector - PostgreSQL extension for vector operations
  • duckdb_vector - DuckDB with vector extension
  • sqlite_vec - SQLite with vector extension

​
Embedding Generation

Generate embeddings automatically using:
  • AWS Bedrock - Amazon Titan, Cohere embeddings
  • HuggingFace - Open-source embedding models
  • Model2Vec - 500x faster static embeddings
  • OpenAI - OpenAI embedding models

​
Distance Metrics

Supported vector distance metrics:
  • Cosine Similarity - Normalized dot product (default)
  • Euclidean Distance - L2 distance
  • Dot Product - Raw inner product

​
Special Columns

Search queries return special columns:
  • _score - Relevance score (0.0 to 1.0 for vectors, float for text)
  • _value - The matched content from the search column
  • _match - Specific substring match (for chunked searches)

​
Architecture

Spice search is built on:
  • Apache DataFusion - SQL query engine and execution
  • Apache Arrow - Columnar data format for zero-copy operations
  • Tantivy - Full-text search library (BM25)
  • Amazon S3 Vectors - Distributed vector storage

​
Use Cases

​
Retrieval-Augmented Generation (RAG)

Search for relevant context to ground LLM responses: 
SELECT content, _score
FROM vector_search(knowledge_base, 'explain quantum computing')
LIMIT 3;
Find documents by meaning, not just keywords: 
SELECH title, summary
FROM vector_search(documents, 'natural language processing applications')
WHERE category = 'AI'
ORDER BY _score DESC;
Search across multiple embedded columns: 
-- Search both titles and descriptions
SELECT id, title, description, _score
FROM vector_search(products, 'wireless headphones', description)
LIMIT 10;

​
Getting Started

  1. Configure a dataset with embeddings
  2. Enable search indexing
  3. Query using vector_search() or text_search() UDTFs
  4. Combine with standard SQL for filtering and sorting
See the individual search method pages for detailed configuration and examples.