Skip to main content

AI capabilities

Databricks provides a comprehensive AI platform for building, deploying, and integrating intelligent applications. This page covers agent frameworks, AI functions, vector search, and integration patterns for partners.

For governance principles, see Data and AI governance for the data lakehouse.

AI agent integrations​

AI agents enable intelligent, context-aware workflows for partner integrations. Partners can build code-based agents or use Databricks AI/BI Genie. Databricks provides a unified, governed foundation for building, running, and integrating agents without maintaining external infrastructure.

AI Agent Framework and MLflow​

Partners can build agents using open-source frameworks like LangChain, LangGraph, LlamaIndex, DSPy, CrewAI or other. The AI Agent Framework provides structured tool invocation, SQL execution, vector search integration, and multi-step reasoning within a Python development model.

MLflow provides the production lifecycle—experiment tracking, packaging, version control, and deployment. Agents run directly on Databricks with low-latency access to governed data, models, vector stores, and tools. Unity Catalog enforces governance across all agent resources.

Documentation: Author AI agents in code | MLflow 3 for GenAI | Deploy models using Mosaic AI Model Serving

Natural language analytics​

Databricks Genie provides a governed natural-language interface for querying Unity Catalog data. Partners configure a Genie Space with curated tables, metadata, and example queries. Genie translates natural-language questions into executable SQL and returns results.

AI agent products and chat-based interfaces must use the Genie APIs via MCP for natural-language data access. Non-agent applications may also use the Genie API directly. Integrations using non-Genie APIs will not be validated.

Documentation: Use Genie in multi-agent systems | Use the Genie API to integrate Genie into your applications

Attribution requirement — Integrations that surface Genie results must display a "Powered by Genie" label and include citations linking to the source Genie Space. See Genie Attribution.

External agents​

Partners with AI agent products running outside Databricks must integrate with Databricks Genie APIs via MCP for platform validation. Foundation models are available via Unity AI Gateway for model inference.

MCP​

Through MCP, external agents can securely delegate tasks to Databricks—executing SQL, retrieving insights via Genie, applying AI Functions, or invoking tools. This supports hybrid architectures where partner agents run externally but rely on Databricks for secure data access and task execution.

Documentation: MCP on Databricks | Unity Catalog Tools

MCP Marketplace​

Partners can publish their own MCP servers to the Databricks Marketplace, making partner tools discoverable and installable by joint customers. Published MCP servers integrate directly with Databricks-hosted agents, enabling tool invocation through a standardized interface.

Documentation: MCP Marketplace Validation

Foundation models​

External agents can access open-source and proprietary models (OpenAI, Anthropic, Google) through a unified serving layer with REST and OpenAI-compatible APIs. No separate infrastructure or credentials needed per provider.

Unity AI Gateway is the Databricks central AI governance layer for agents, LLM endpoints, MCP servers, and coding agents. Use Unity AI Gateway to analyze usage, configure permissions, enforce guardrails, and manage capacity across providers.

Documentation: Foundation Models | Foundation Model APIs | Unity AI Gateway

Databricks-hosted agents and LLMs​

Partners can invoke Databricks-hosted agents, LLMs, and custom models through several fully supported interfaces that provide flexibility to integrate. These interfaces are for querying Databricks-deployed models from any application. AI agent products must also satisfy the Managed MCP integration requirement for data access, tool invocation, and Genie workflows.

  • AI Functions, such as ai_query, callable from SQL or Python to access foundation models, Databricks-hosted models, agents, or external models through a single governed interface. They are also optimized to support batch inference use cases.
  • REST APIs, which allow any external application or service to send inference requests to Databricks-deployed models or agents.
  • OpenAI-compatible APIs, enabling partners to use standard OpenAI client libraries and request formats to call Databricks models and agents with minimal code changes.

Documentation: Model Serving | AI Functions

AI functions​

AI Functions are built-in SQL functions for applying AI to data stored on Databricks. They run from Databricks SQL, notebooks, Lakeflow SDP, and Workflows.

  • ai_query - General-purpose function to call any AI model (Foundation Model APIs, external models, custom ML models) through a single SQL interface
  • Task-specific functions - High-level capabilities like summarization and translation, powered by Databricks-managed models

Use cases​

  • BI tools - Summarize dashboards, explain anomalies, generate natural-language insights
  • ETL pipelines - Extract entities, summarize documents, generate structured fields from unstructured data
  • SQL-based tools - Add text generation or reasoning capabilities where the data lives

Documentation: AI Functions | ai_query | Task-specific Functions

Databricks Vector Search is a serverless, production-grade vector database built into the platform. Create and query vector search indexes backed by Delta tables, governed by Unity Catalog.

For ingest partners​

Compute embeddings and write them to Delta tables, then create indexes using:

  • Delta Sync Indexes - Automatically sync vectors from a source Delta table on updates
  • Direct Vector Access (DVA) - Upsert embeddings directly via the Vector Search API

Use your own embedding models or Databricks-hosted models. All indexes are registered in Unity Catalog.

For agent partners​

AI agent products should query Vector Search through Databricks Managed MCP, which exposes Vector Search as a governed tool. Direct REST API or SDK access remains available for non-agent integrations.

For non-agent integrations, query indexes directly via REST API or SDK:

  • similarity_search - Nearest neighbors by vector similarity
  • query with metadata filters - Match embeddings and structured filters

Documentation: Vector Search | Create and Query Indexes | Managed MCP | Agent Bricks

What's next​