Databricks Mosaic AI

Mosaic AI is Databricks' generative AI platform — the integrated set of services for building, evaluating, serving, and governing GenAI applications directly on the Lakehouse. Rather than stitching together a vector database, an LLM gateway, an evaluation harness, and a separate orchestration framework, Mosaic AI ships these as cohesive products that share Unity Catalog identity, lineage, and access control.

The core building blocks are:

• Vector Search — managed similarity search built on Delta tables, with automatic sync from source data and UC-governed retrieval.
• Model Serving — low-latency REST endpoints for foundation models, fine-tuned models, and custom Python models, with autoscaling and request logging into Inference Tables.
• Foundation Model APIs — pay-per-token access to Llama 3, DBRX, Mixtral, Claude, and embedding models without provisioning capacity.
• Agent Framework — an opinionated SDK and evaluation harness for building, tracing, and shipping LLM agents that run as production endpoints.
• AI Playground — a workspace-native chat UI for prompt iteration, model comparison, and quick prototyping that exports straight to notebook code.
• AI Functions — SQL functions like ai_query, ai_classify, and ai_summarize that call serving endpoints from inside warehouse queries.

Coming next

Natural extensions to this section, planned as follow-on pages:

• Agent Framework — building tool-calling agents with mlflow.deployments, ChatAgent, and the Agent Evaluation harness; deploying as a Mosaic AI Agent endpoint.
• Model Serving — provisioned-throughput vs pay-per-token, custom Python model deployment, Inference Tables for request logging, and traffic splitting for A/B tests.

1. How the Pieces Fit

The services in Mosaic AI share one idea: generative AI components should live next to the governed data they use. Source documents sit in Delta tables, a Vector Search index stays in sync with a table through Delta change data feed, served models and agents run behind Model Serving endpoints, and Unity Catalog applies the same permissions and lineage to tables, indexes, models, and functions. Vector Search is the page in this section today and is the usual starting point, because retrieval quality drives most RAG results.

Featured Videos