Skip to main content
Version: Next

Spice.ai Features

Spice provides a set of features for building data-driven applications and AI agents. This page gives an overview of each feature area.

Data Query and Federation​

Query Federation connects multiple data sources—databases, data warehouses, and data lakes—through a single SQL interface. Write one query that joins data across PostgreSQL, Snowflake, S3, and other sources. Spice pushes query operations to source databases when possible to reduce data transfer.

Data Acceleration and Caching​

Data Acceleration materializes remote datasets locally in memory or on disk using engines like Arrow, DuckDB, SQLite, or PostgreSQL. Accelerated datasets stay current through scheduled refreshes, append mode, or Change Data Capture (CDC). Caching stores query and search results in memory with configurable TTLs and eviction policies to avoid redundant computation.

Views​

Views create virtual tables from SQL queries over other datasets, similar to database views — useful for encapsulating query logic and (when accelerated) materializing precomputed joins or aggregates.

AI and Language Models​

Large Language Models provides an OpenAI-compatible API gateway for hosted models (OpenAI, Anthropic, xAI) and locally served models (Llama, Phi) with CUDA and Metal acceleration. Models can call tools to query datasets, run SQL, and retrieve schemas. Embeddings generates vector representations of text for semantic search and RAG workflows. Workers coordinate interactions between models and tools, supporting load-balancing strategies such as round-robin and fallback across multiple LLM providers.

Search supports three methods: vector search (semantic similarity using embeddings), full-text search (keyword matching with BM25 scoring), and hybrid search (combining both with Reciprocal Rank Fusion). All search methods are accessible through SQL UDTFs like vector_search() and text_search().

Functions​

Functions extend SQL with custom scalar functions declared in a Spicepod. Inline SQL bodies run in-process and can use any DataFusion built-in; remote http:// / https:// endpoints batch row inputs over JSON for delegating logic to ML models, internal services, or custom code. Every function is automatically callable from SQL and (by default) surfaced as an LLM tool.

Tool Registry​

Tool Registry keeps per-turn token cost bounded as the runtime's tool catalog grows. It replaces individual tool definitions with searchable tool_search and tool_invoke meta-tools backed by a hybrid full-text, keyword, schema, and vector search. Applies uniformly to built-in tools, MCP tools, and Functions declared with as_tool: true — typically a ~10× reduction in tool-definition tokens for tool-heavy Spicepods.

Monitoring and Observability​

Observability exposes Prometheus-compatible metrics, OpenTelemetry metric export, and distributed tracing with Zipkin. Integrations are available for Datadog, Grafana, and other monitoring platforms.