Databricks Mosaic AI

Mosaic AI is the group of Databricks tools for building, deploying, and managing artificial intelligence applications. The tools cover classic machine learning and generative AI under one roof. Teams keep their data, models, and applications in the same governed platform, so an AI project does not need a separate set of systems.

The Toolbox Analogy

A carpenter owns a toolbox, and each tool has a clear job. A saw cuts, a drill makes holes, and a level checks the result. Mosaic AI works like a toolbox for AI builders. One tool serves language models, one tool finds related documents, one tool builds assistants, and one tool measures quality.

THE MOSAIC AI TOOLBOX
──────────────────────────────────────────────────────────
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ FOUNDATION      │ │ VECTOR SEARCH   │ │ AGENT           │
│ MODEL APIS      │ │ Find related    │ │ FRAMEWORK       │
│ Use language    │ │ text and data   │ │ Build assistants│
│ models          │ │                 │ │ that use tools  │
└─────────────────┘ └─────────────────┘ └─────────────────┘
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ MODEL SERVING   │ │ AI FUNCTIONS    │ │ EVALUATION      │
│ Host models as  │ │ Call AI from    │ │ AND TRACING     │
│ live endpoints  │ │ SQL             │ │ Measure quality │
└─────────────────┘ └─────────────────┘ └─────────────────┘
          All governed by Unity Catalog

Main Components

Foundation Model APIs

Databricks hosts large language models and exposes them through simple endpoints. Applications send text and receive generated text without managing any servers.

Vector Search

Vector Search stores numeric fingerprints of text, called embeddings, and finds the items most similar to a question. It powers search by meaning instead of search by exact words.

Agent Framework

The framework helps developers build assistants that reason through a request, call tools, and return an answer. Developers log, test, and deploy agents with MLflow.

AI Playground

The playground is a chat window inside the workspace. Teams try prompts, compare models, and adjust settings before they write any code.

AI Functions

AI Functions bring language models into SQL. An analyst summarizes thousands of customer comments with one query.

SELECT review_id,
       ai_summarize(review_text)         AS short_summary,
       ai_analyze_sentiment(review_text) AS sentiment
FROM main.shop.product_reviews
LIMIT 20;

The query reads each review, asks a model to summarize it, and labels the mood as positive, negative, neutral, or mixed.

Classic Machine Learning and Generative AI

Traditional models predict numbers and categories, such as customer churn or delivery time. Generative models create text, images, or code. Mosaic AI supports both styles. A single project often combines them: a churn model flags customers at risk, and a language model drafts a helpful message for the support team.

Style	Typical Task	Common Tools
Classic ML	Predict churn, price, or demand	AutoML, MLflow, Feature Store, Model Serving
Generative AI	Summarize, answer, draft, extract	Foundation Model APIs, Vector Search, Agents

The Lifecycle of an AI Application

AI APPLICATION LIFECYCLE
──────────────────────────────────────────────────────────
Prepare data ──► Choose a model ──► Build the app ──► Evaluate
      ▲                                                  │
      │                                                  ▼
Monitor and improve ◄──────────── Deploy ◄───────────────┘

Each stage has matching tools. Unity Catalog tables hold the data. Foundation Model APIs supply the model. The Agent Framework builds the logic. Evaluation tools score the answers. Model Serving deploys the result, and monitoring tools watch it after launch.

Governance Built In

Unity Catalog applies the same permissions to AI assets as it does to tables. Models, functions, and vector indexes carry owners and access rules. Lineage shows which data feeds which model. Audit logs record who called an endpoint. Companies in regulated industries rely on these records to answer questions about how an AI system works.

Example: Customer Support Helper

An electronics retailer wants faster answers for its support agents. Engineers store product manuals in a table and index them with Vector Search. An agent searches the manuals, reads the best passages, and drafts a reply. A human agent reviews the draft before it reaches the customer.

SUPPORT HELPER
──────────────────────────────────────────────────────────
Customer question
      │
      ▼
Agent ──► Vector Search finds manual pages
      │
      ▼
Foundation model writes a draft answer
      │
      ▼
Support agent reviews ──► sends reply

Responsible AI Habits

• Test answers on many sample questions before launch.
• Keep a human in the loop for decisions that affect people.
• Remove private personal data from prompts and training sets.
• Tell users when they interact with an AI system.
• Log requests so teams can review mistakes and improve.

Key Points

• Mosaic AI groups the Databricks tools for classic machine learning and generative AI.
• Foundation Model APIs, Vector Search, agents, and AI Functions cover common building tasks.
• AI Playground lets teams test prompts without code.
• Unity Catalog governs AI assets with the same rules as data.
• Evaluation and monitoring keep applications reliable after launch.
• Human review and clear disclosure support responsible use.