Genie & Conversational Analytics

Genie lets business users ask data questions in natural language, and the docs state that every answer is grounded in your organization's data and governed through Unity Catalog. Data teams create a trusted data environment by configuring datasets, sample queries, and instructions. Genie is also available as a mobile app on iOS and Android.

Yes. The Genie Conversation API provides REST access to Genie in Chat mode, for stateful follow-up conversations, and in Agent mode, and it can also be called through the Databricks SDKs. Applications can retrieve the generated SQL, query results and visualizations, and list and delete an agent's conversations through the management APIs. Responses are asynchronous, so the documentation recommends polling until the status reaches a terminal state such as COMPLETED.

Genie Code is Databricks' AI coding and data assistant, available inside the workspace across notebooks, the SQL editor, pipelines, jobs, and AI/BI dashboards. It provides inline suggestions and runs agentic tasks, such as generating and running code, building pipelines and dashboards, and suggesting fixes for errors.

Genie One is a simplified experience that lets business users find and query data assets in one place. It brings together dashboards, Genie's natural-language data questions, and Databricks Apps, and it is available on mobile devices. Because Genie answers are grounded in data governed through Unity Catalog, business users can get trusted answers without writing SQL.

The Genie Conversation API provides REST access to Genie's natural-language capabilities in Chat mode and Agent mode, and it can also be called through the Databricks SDKs. Applications can retrieve the generated SQL, query results, and visualizations, and they can manage conversation history through the management APIs. Because responses are asynchronous, the documentation recommends polling status until it reaches a terminal state such as COMPLETED.

Unity Gateway governs access to MCP servers that agents use, alongside models and skills, with guardrails, rate limits, and usage tracking. For Genie, a workspace admin enables Managed MCP Servers on the Previews page to expose a Genie space as an MCP server, and Genie Agents can also connect to external tools through MCP connectors.

Unity Gateway governs access to MCP servers alongside foundation models and skills, applying guardrails, rate limits, and usage tracking to them. With MLflow Tracing, you can store and govern traces of agent steps at scale in Unity Catalog for review.

Yes. Genie One lets business users find and work with governed data assets through natural-language questions, with answers grounded in your organization's data through Unity Catalog. Data teams define trusted data, metrics, and business rules in Genie Agents, which power the answers that Genie One returns. Genie One is also available on iOS and Android.

Analysts query governed data with Databricks SQL and ask questions in natural language through Genie One, with answers grounded in data that Unity Catalog permissions govern. Data teams can define trusted metrics and business rules in Genie Agents to power consistent answers.

Genie One answers are grounded in your organization's data through Unity Catalog, so they reflect the same governed data your teams already use. Data teams set up trusted data, metrics, and business rules in Genie Agents, and those power the answers that business users receive. Genie is also available on iOS and Android.

Yes. Genie Code is an AI coding and data assistant that helps developers across notebooks, pipelines, and dashboards, working with governed data in Unity Catalog. Describe what you want, review the code and assets it proposes, and refine them before you deploy. Genie Code moved to pay-as-you-go billing on July 8, 2026, with a per-user free monthly allowance.

Genie Agents is the environment where data teams set up trusted data, metrics, and business rules that power answers in Genie One. Business users ask questions in natural language and receive answers grounded in governed data, with Unity Catalog controlling access. Genie is also available on iOS and Android.

Unity Gateway is the Databricks governance solution for enterprise AI, built on Unity Catalog. It governs models, agents, MCP servers, and tools, with audit logging written to Unity Catalog Delta tables. Use the Genie Conversation API to embed Genie in custom chatbots, agents, and applications.

Business users ask Genie questions in plain language. Creators improve answers with table and column descriptions, primary and foreign key relationships (which are saved as join relationships automatically), example SQL queries, and general text instructions. Users with CAN EDIT on an agent can view the query used to generate a response.

The Genie Conversation API is part of Genie API chat mode and is used to build natural-language data querying into custom chatbots, agents, applications, and agent frameworks. Conversations are stateful, so follow-up questions keep context. Databricks advises a new conversation for each user session, because reusing a thread across sessions can reduce accuracy. The docs do not state a release status, and do not list Slack, Teams, or SharePoint integrations.

Databricks Assistant autocomplete is documented under Genie Code as AI-based autocomplete in Databricks notebooks, the SQL editor, and the file editor. Suggestions appear as you type, and you press Tab to accept one. The docs do not state a GA or Beta status, so none is asserted.

Genie answers natural-language questions over governed data in Unity Catalog and returns text summaries, tables, and visualizations. Genie Agents were formerly called Genie Spaces. The docs page does not state a general availability date.

The Supervisor Agent coordinates Genie Agents, agent endpoints, Unity Catalog functions, MCP servers and custom agents. It is a legacy offering and is not recommended for new systems; Databricks recommends Genie Agents for low-code agents and the Agent Bricks CLI for code-based multi-agent orchestration.

Lakehouse Federation lets Databricks query external sources such as Snowflake and other listed systems through foreign catalogs without copying the data, and Genie Agents answer natural-language questions over governed data. The manufacturing scenario and the traceability example come from a Databricks blog.

Genie One is a simplified interface for business users to find and interact with data assets, including shared assets across workspaces. Genie Agents are domain-specific environments where data teams configure datasets, sample queries, and instructions. The docs do not describe Genie Ontology, and do not give GA status for Genie One or Genie Agents.

Genie Code is an AI coding and data assistant in the Databricks workspace that covers notebooks, pipelines, and dashboards, with custom instructions at the workspace and personal levels. Lakeflow Designer and Lakeflow Connect have their own docs and are not covered on this page.

Genie Code's agentic code converter converts T-SQL, Snowflake, Redshift, Oracle, BigQuery and Teradata SQL into ANSI-compliant Databricks SQL. The SQL target became generally available on October 15, 2026, while the notebooks (DataFrame API) target is in Beta and requires a workspace admin to enable a preview.

Agent mode, formerly Research Agent, answers complex exploratory questions in a Genie Agent using multi-step reasoning and hypothesis testing. It creates a research plan, runs several SQL queries, and produces a report with citations, visualizations, and supporting tables. Flash mode is in Beta and must be turned on by a workspace admin on the Previews page.

Agent mode, formerly Research Agent, uses multi-step reasoning and hypothesis testing to answer complex, exploratory questions in a Genie Agent. It creates a research plan, runs several SQL queries, and returns a report with citations, visualizations, and supporting tables. Flash mode is Beta and needs a workspace admin to enable Genie Agents Flash Mode on the Previews page; Inspect is Public Preview.

The Databricks-provided MCP servers include a Genie One MCP server, which the docs recommend for natural-language business questions. Answers are grounded in Genie Ontology, and registered MCPs use Unity Catalog permissions, with Unity Gateway used to control access, manage credentials, and monitor usage. The docs do not label the Genie One MCP as GA.

Genie Code is an AI coding and data assistant inside the Databricks workspace that covers notebooks, pipelines, and dashboards. Its pipeline failure diffs and bronze, silver, and gold pipeline generation are not described on the docs page.

Databricks Apps runs on serverless compute and can host data and AI apps that authenticate users through SSO. An app can attach a Genie Agent or a model serving endpoint as an app resource, and access to each resource is set through the app's service principal or, for user-authorized calls, the user's Unity Catalog permissions.

Genie answers depend on business context, and the docs recommend supplying table and column descriptions, join relationships (primary and foreign keys are saved as join relationships automatically), example SQL, SQL expressions for KPIs and business terms, and general instructions. Genie operates in a nondeterministic manner, so ambiguous or conflicting guidance can produce undesirable responses.

Genie Code is an AI coding and data assistant inside the Databricks workspace. It covers notebooks, pipelines, and dashboards, and it supports custom instructions at the workspace and personal levels.

The Genie docs describe Genie One, a simplified interface for business users to find and interact with data assets; Genie Agents, domain-specific environments where data teams configure datasets, sample queries, and instructions; and Genie Code, an AI coding and data assistant in the workspace.

Declarative Automation Bundles define Lakeflow jobs, Lakeflow pipelines, and dashboards as code in YAML, which can be version-controlled and promoted across environments. The bundle docs do not describe Genie spaces or metric views as bundle resources, so the Genie-as-code workflow and its CLI requirement are not confirmed.

S&P Global Energy curated a Genie Agent per dataset group, with subject-matter experts choosing tables and adding descriptions, example queries and business definitions without writing agent code. Each Genie Agent is exposed as a managed MCP server governed by Unity Catalog, and the post reports qualitative results rather than measured figures.

Knowledge Assistant builds a Q&A agent over up to 10 knowledge sources, which can be Unity Catalog volume files, a Unity Catalog table, or an AI Search index. Sync after changing source files so updates are picked up, and add expert feedback as guidelines. The docs label Knowledge Assistant a legacy offering, and Genie Agents is recommended for new low-code agents.

Genie Ontology is the unified context layer for the Genie family of AI experiences, combining modeled context (Unity Catalog semantics that you define and certify) with inferred context that Genie extracts from existing assets and usage. It is in Public Preview.

Set up a Genie Agent (formerly Genie Spaces) over up to 50 trusted Unity Catalog tables, views, or metric views so business users can ask natural-language questions and get answers without writing SQL. Genie Code suggests table descriptions and example queries that help the agent answer accurately. Genie Agents require a pro or serverless SQL warehouse. This gives teams self-service access to governed data.

The Databricks SQL editor lets you run single or multi-statement queries against a SQL warehouse, browse catalogs and schemas, and author SQL with autocomplete and Genie Code. You can share queries with collaborators and edit them together in real time, and you can visualize and filter results. The editor keeps a version history of your query so you can review recent changes.

The Supervisor Agent docs describe a supervisor that coordinates Genie Agents, agent endpoints, Unity Catalog functions, MCP servers and custom agents, and end users reach only the subagents and data they have permissions for. The page labels Supervisor Agent a legacy offering and recommends Genie Agents or the Agent Bricks CLI for new systems.

Agent mode takes one natural-language question, builds an investigation plan with hypotheses and data to examine, runs several SQL queries, and produces a report with citations, visualizations, and supporting tables. Flash mode is Beta and needs admin enablement on the Previews page.

The system.data_quality_monitoring.table_results system table stores freshness and completeness check results for tables enabled for data quality monitoring in your metastore, with a consolidated status of Healthy, Unhealthy or Unknown. The feature is in Public Preview, and only account admins can read it by default.

Unity Gateway governs MCP servers registered as Unity Catalog securables, with tool filtering and service policies that add guardrails based on request and response content and on the caller. Request and response logs can be written to Unity Catalog Delta tables for monitoring and debugging.

Yes. Agent memory lets agents recall what they learned in earlier conversations to personalize responses and reduce token costs. Conversation state and long-term memory are stored in managed stores backed by Lakebase. The docs do not give a release status for this feature.

Agent memory stores conversation state and long-term memory in managed stores backed by Lakebase, so agents can recall what they learned in earlier conversations. The docs do not describe the MemCurrent architecture or its benchmark results.

Agent memory can reduce token costs by letting agents recall what they learned in earlier conversations, and the docs say so as a goal. The docs do not quantify the savings, so measure token use on your own tasks.

Agent memory stores conversation state and long-term memory in managed stores backed by Lakebase, so agents can recall what they learned in earlier conversations. The docs do not describe per-user feedback adaptation or retraining behavior, so validate it on your own workflows.

Agent memory can reduce token costs by letting agents recall prior learnings from earlier conversations. The docs do not quantify response-time gains.

Genie is current. Databricks describes it as the AI experience for users, a single place to ask data questions in natural language, explore AI/BI dashboards, and run Databricks Apps. It now includes Genie Agents, where data teams set up trusted data, metrics, and business rules, and Genie Code, an AI coding and data assistant for developers.

Genie Agents are spaces where data teams configure the trusted data, metrics, and business rules that power Genie One answers, so responses reflect how your organization defines its numbers. Answers are grounded in your organization's data and governed through Unity Catalog. This lets a data team curate context for business users rather than leaving every question to raw table exploration.

Genie answers are grounded in your organization's data and governed through Unity Catalog, but Databricks makes no guarantee that they are correct. Genie Agents let your data team define trusted data, metrics, and business rules that power answers. Benchmark the agent with questions whose answers you know, and review results before using them for high-stakes decisions.

Genie Code uses pay-as-you-go billing, and each identified user gets a free monthly allowance of LLM usage. Usage beyond that allowance is billed in DBUs, beginning July 8, 2026 at 00:00 UTC. Service principals do not get the free allowance and are billed for all of their usage. Budgets can alert on or block pay-as-you-go spend, which gives you direct control over the billable portion.

Genie budgets track the LLM side of usage through Unity Gateway, billed in DBUs based on underlying LLM usage after each user's free monthly allowance. Compute used for queries, such as a SQL warehouse, is billed separately and falls outside Genie budget tracking. Budget the two as separate cost lines to get clear visibility into each.

Genie budgets cover LLM usage through Unity Gateway, and compute used for queries, such as a SQL warehouse, is billed separately. The authoritative record of billable usage is system.billing.usage, which updates every few hours. Use it alongside Genie budget alerts to reconcile total spend.

Genie Code is billed pay-as-you-go with a per-user free monthly allowance. Usage by users of Genie One and Genie Agents is free through January 31, 2027. Service principals are excluded from that promotion and are billed for their usage, and Genie compute such as a SQL warehouse is billed separately from Genie usage.

Databricks recommends starting with minimal instructions and a limited set of questions, then expanding based on feedback and monitoring. Prefer SQL expressions for business terms such as metrics and filters, and example SQL for complex questions, because structured definitions are more reliable than plain text. Use text instructions as a last resort when SQL cannot express the need, and keep them consistent with example SQL, such as a rounding rule that matches the SQL.

Keep the agent focused, ideally with five or fewer tables and a limited number of columns, and hide columns that could confuse Genie. Write clear Unity Catalog table and column descriptions, which Databricks calls critical for Genie accuracy, and add column synonyms where useful. Add complete example SQL for complex or ambiguous questions, then test the agent yourself and save reviewed questions as benchmarks.

Keep the set of tables small and focused, ideally five or fewer tables with a limited number of columns. A Genie Agent accepts up to 50 tables, views, or metric views, and if a topic needs more objects, pre-join tables into views or metric views. Metric views pre-define metrics, dimensions, and aggregations, which simplifies the data model and can improve accuracy.

Databricks recommends starting as small as possible, with minimal instructions and a limited set of questions the agent needs to answer. Begin with a focused set of tables that have clear Unity Catalog descriptions rather than modeling everything up front. Expand with example SQL, SQL expressions for business terms, and metric views as real questions show where the agent needs more context.

Clear Unity Catalog table and column descriptions are a critical foundation for Genie accuracy, and Databricks recommends them as the starting point. Pair them with column synonyms, example SQL for ambiguous prompts, SQL expressions for business terms, and metric views that pre-define metrics. Then check results against a set of benchmark questions before rolling the agent out.

Metric views pre-define metrics, dimensions, and aggregations in one place, so each measure is defined once and reused. They simplify the data model and can improve accuracy for Genie. Metric views also help keep large topics within the 50-object limit of a Genie Agent, and you can pre-join tables into them so Genie queries one well-defined object.

When you create a Genie Agent, Genie Code launches automatically and can suggest context such as table descriptions and example queries, or build the agent for you. An existing agent can be exported to a metric view that captures its configured data and semantic context. The resulting metric view defines measures and dimensions in one place for reuse.

Genie Ontology, currently in Public Preview, has Genie Code search your organization's business context, such as metric definitions and join hints, before it writes SQL. Grounding SQL generation in your metric definitions and join paths means the generated queries follow how your business defines terms. Databricks does not publish a measured accuracy figure, so validate with benchmarks built from your own questions.

Genie Code sends your prompt plus relevant context, including table and column names and their descriptions, the code you are working with, data samples from tables, and cell outputs. Its access is governed by your Unity Catalog permissions, so it only uses data you are allowed to see. The docs list these context sources explicitly, so you know what goes to the model.

Use Genie Agents for natural-language questions over structured data registered in Unity Catalog. Knowledge Assistant answers questions over documents from up to 10 sources, such as files in a Unity Catalog volume, a table with a content column, or an AI Search index. Databricks labels Knowledge Assistant a legacy offering and recommends Genie Agents for new low-code agents, so start new work there.

Databricks documents the Genie Conversation API for integrating Genie into applications and chatbots. Start a new conversation for each session, because reusing threads across sessions can lower accuracy through unintended context reuse. Poll for status every 1 to 5 seconds with exponential backoff up to one minute, and stop after 10 minutes. Use OAuth for users in production, or a service principal when browser-based authentication is not possible.

The Conversation API lists the conversation threads for a space through GET /api/2.0/genie/spaces/{space_id}/conversations, which lets you pull chats for analysis. The docs ask you to stay under the 200,000-conversation limit, so delete unneeded conversations with the DELETE conversation endpoint. Keep polling until a conclusive status such as COMPLETED, FAILED, or CANCELLED is returned.

Write clear, specific instructions in plain English, grouped under Markdown headings with bullets for individual rules. Include preferred libraries, coding conventions, role context, and response style, along with details Genie Code cannot infer, such as which tables or functions to use and when. Genie Code does not proactively fetch extra information based on instructions, so name what it needs. Keep each file under 20,000 characters, because Genie Code ignores anything past that limit.

Genie Code is an autonomous agent: it plans solutions, runs code, uses cell outputs to refine results, and fixes errors, and it works within your Unity Catalog permissions. It can also answer conversationally without executing code. The docs reviewed do not describe default approval behavior for interactive sessions; scheduled tasks auto-approve tool actions, with an AI classifier blocking risky ones.

Genie Code is the AI coding and data assistant for developers and technical practitioners in the Databricks workspace. It generates and runs code, builds pipelines and AI/BI dashboards, and debugs errors across notebooks, the SQL editor, jobs, and dashboards. Databricks positions Genie One for business users, Genie Agents for data teams that configure trusted data, and Genie Code for developers.

Create benchmarks, which are test questions you run against the agent, and each agent can hold up to 500. In Chat mode, Genie compares the result set of your example SQL with the SQL it generates, and questions without a SQL answer need manual review. In Agent mode, an LLM judge grades responses using an optional evaluation note. Databricks recommends two to four phrasings of each question, and each run's accuracy appears on the Evaluations tab.

Start small with a focused set of tables and minimal instructions, which is the approach Databricks recommends for a great Genie experience. Write clear Unity Catalog descriptions, prefer SQL expressions for business terms, and add complete example SQL for complex or ambiguous questions. Pre-define metrics with metric views where possible, then test the agent yourself and save reviewed questions as benchmarks, keeping example SQL consistent with any text instructions.

Clone a Genie Agent to create an independent copy that includes its tables, settings, instructions, example SQL queries, and SQL functions, but not chat threads or monitoring data. Share the agent with specific users or groups, or with all account users, at CAN MANAGE, CAN EDIT, CAN RUN, or CAN VIEW. You can also mark agents as Certified or Deprecated with the system.certification_status governed tag, which requires the ASSIGN permission on that tag.

Business users ask natural-language questions in a Genie Agent and receive the SQL, result tables, and visualizations. Data analysts curate each agent with Unity Catalog datasets, example SQL queries, SQL expressions for business terms, and text instructions that match the organization's vocabulary. This gives business users self-service analytics grounded in governed data.

You do not have to build every piece by hand. When you create a Genie Agent, Genie Code launches automatically and suggests table descriptions and example queries, which you can review and accept. You can also ask Genie Code to build the agent for you. Before sharing, add example SQL, instructions, and knowledge store configurations to improve accuracy.

Creating a Genie Agent requires the Databricks SQL workspace entitlement, CAN USE on at least one pro or serverless SQL warehouse, and SELECT privileges on the data the agent uses. Creators automatically receive CAN MANAGE on the agent, and users who configure it need at least CAN EDIT. The agent is built from Unity Catalog data you already govern, so access stays under your existing controls.

A Genie Agent supports up to 50 tables, views, or metric views, each registered in Unity Catalog. Databricks recommends starting with five or fewer tables and limiting the columns in each one, which keeps answers focused. Pre-joining related tables into views or metric views lets Genie query one well-defined object, and sources are managed under the agent's data configuration.

Genie Agents run on a pro or serverless SQL warehouse. Saving a warehouse embeds the configuring user's compute credentials for all users, so end users don't need their own warehouse permissions. Data access is still checked against each end user's Unity Catalog permissions, so governance stays per user.

If the author who last configured the agent's SQL warehouse is removed, their embedded credentials stop working and queries fail for everyone using the agent. The fix is quick: a user with at least CAN EDIT reconfigures the agent's SQL warehouse, which embeds that user's credentials instead. Assigning a CAN EDIT owner as part of offboarding keeps agents running without interruption.

Yes. Genie Code launches automatically when you create an agent and suggests table descriptions and example queries that you can accept or reject. You can also describe the business domain and have it identify relevant data for the agent.

From the Share dialog you can add users or groups with CAN MANAGE, CAN EDIT, CAN RUN, or CAN VIEW, or select All account users to share broadly. Agents inherit permissions from their workspace folder, and the Copy link button gives you a direct link to share. Recipients need consumer access or the Databricks SQL entitlement, plus SELECT on the underlying data.

A workspace admin configures the allowed embedding surfaces in the Genie access management settings. An agent author with at least CAN MANAGE then opens the agent, selects Share, then Embed space, and copies the iframe code. Adding the allow=clipboard-write attribute lets embedded users copy CSV data and conversation links, and embedded users can send prompts to the agent.

A Genie Agent works with managed tables, external tables, foreign tables, views, metric views, and materialized views registered in Unity Catalog. It relies on the metadata attached to these objects together with the agent's knowledge store. Agent-level synonyms and descriptions apply only to the agent and do not change Unity Catalog metadata, so you can tune the agent without changing shared catalog definitions.

Databricks recommends prejoining related tables into views or metric views so Genie queries one well-defined object instead of inferring joins. Metric views pre-define metrics, dimensions, and aggregations. The troubleshooting guide also lists a pre-joined view as one of the fixes when Genie joins tables incorrectly.

Yes. Column synonyms and custom descriptions are added inside the agent, and that metadata stays scoped to the agent without overwriting Unity Catalog. Databricks still recommends writing clear column names and precise descriptions in Unity Catalog, and inspecting AI-generated descriptions for accuracy.

Prefer SQL expressions for reusable metrics, filters, and dimensions, and example SQL for complex, multi-part, or ambiguous questions. Use text instructions only when SQL cannot express the need, such as rounding rules or clarification behavior. Keep all instruction types consistent, so text rounding rules match the rounding in example SQL.

Each Genie Agent supports up to 100 instructions. Each example SQL query, each SQL function, and the general text instructions block each count as one instruction. Agents also support up to 200 knowledge store snippets, which cover table descriptions, join relationships, and SQL expressions, so budget your context toward the most valuable items.

Entity matching supports up to 120 columns, with up to 1,024 distinct values per column, and each value can be up to 127 characters. It applies to string columns, and authors can manage which columns have prompt matching enabled. These limits keep entity matching focused on the columns where users refer to specific entries.

Trusted assets are parameterized example SQL queries and SQL functions whose logic an agent author has verified. In chat mode, Genie returns a verified answer when the exact text of a parameterized query is used to generate the response. Users with at least CAN EDIT can add or remove trusted assets, and agent users need EXECUTE permission on any SQL functions used.

Users can rate each response as Yes, Fix it, or Request review, and that feedback is visible to users with CAN MANAGE. Genie does not change its behavior automatically from feedback, so managers review it and update instructions, example SQL, or trusted assets to improve the agent. This keeps every change deliberate and reviewed.

Each Genie Agent can contain up to 500 benchmark questions. To add one, open Benchmarks, select Add benchmark, and enter a question with an optional SQL answer that you write or generate with Generate SQL. The docs recommend testing two to four phrasings of the same question to measure accuracy across variations.

Each benchmark run appears on the Evaluations tab with its timestamp, execution status, accuracy, and the user who ran it. Click a run to compare model output with ground truth, update the ground truth when a better answer exists, and mark each result Good or Bad for an accurate score. Detailed results stay visible for one week, and the generated and example SQL remain available after that.

Structure each clarification instruction with four parts: a trigger condition, the missing details, a required action, and an example clarifying question. Place these clarification instructions at the end of your general text instructions. This gives Genie a defined, consistent behavior for ambiguous questions.

Add a section headed Instructions you must follow when providing summaries at the end of your text instructions. Only text instructions affect summary wording, so example SQL and SQL expressions do not change summaries. Keep this section aligned with the rest of your instructions for consistent results.

The Analysis section shows Genie's thinking steps, which explain the SQL logic in plain language. In the result table, click Show code to inspect the generated query. For older responses, click Show generated code and then Run to rerun the query without regenerating the response.

Chat mode works with structured data and is the fast path for data questions, but it cannot analyze files or use MCP connectors. Agent mode handles complex, exploratory questions with multi-step reasoning, and returns detailed summaries with citations, visualizations, and supporting tables. Agent mode takes longer to respond, and it is where file analysis and MCP tools are available.

Flash (Beta) is an Agent mode option for quick data questions. Turn it on by clicking the Flash icon in the chat box before you send the first message in a conversation. Once a conversation has started, Flash can no longer be turned on for it.

Query results persist for seven days, after which you can regenerate the response or rerun the SQL to get them again. Results can be downloaded as CSV, with a download limit of approximately 1GB of data. Visualizations download as PNG, and Agent mode reports export as PDF with findings, visualizations, and citations.

Data access is evaluated against each end user's own identity, so Unity Catalog row filters and column masks are enforced per user. A question about data the user is not authorized to access returns an empty response. One agent can serve many audiences while each person sees only the data they are allowed to see.

Genie cancels SQL queries that run longer than five minutes, and it cancels messages that run longer than 30 minutes. If responses time out, the troubleshooting guide recommends using trusted assets or views to encapsulate complex queries. Moving complex logic into a view or trusted asset keeps the agent's generated SQL simple.

Enable Example values and Value dictionaries for the relevant columns so Genie can see the values in your data, and refresh those values when new data is added. The troubleshooting guide lists these settings as the documented fix for filtering errors. With them enabled, filters are matched against the values that actually exist in your tables.

The troubleshooting guide lists four fixes for incorrect joins: define foreign keys in Unity Catalog when possible, define join relationships in the agent's knowledge store, add example queries that join tables in standard ways, or pre-join the tables into a view and use that as the agent's input. The knowledge store option helps with self-joins or when you lack permission to modify the tables, so start with the option that fits your permissions and data model.

Each Genie Agent supports up to 200,000 conversations, so deleting old or test conversations keeps you under the limit. Each conversation supports up to 10,000 messages. Use the Conversation API to list and delete threads you no longer need.

Poll every one to five seconds and increase the delay with exponential backoff up to one minute. Keep polling until a final status such as COMPLETED, FAILED, or CANCELLED is returned. Limit polling to about 10 minutes for most queries, then return a timeout message or ask the user to check back later.

GET /api/2.0/genie/spaces lists the agents, and returns the space_id you need. GET /api/2.0/genie/spaces/{space_id}/conversations lists the conversation threads in an agent. DELETE /api/2.0/genie/spaces/{space_id}/conversations/{conversation_id} deletes one conversation, which lets you manage conversation volume programmatically.

The workspace needs the Databricks SQL entitlement, and the caller needs at least CAN USE on a pro or serverless SQL warehouse. A service principal also needs access to the required data and SQL warehouses, and it authenticates when browser-based login is not possible. Use OAuth for users who have a browser, and a service principal for non-interactive automation.

Content search supports PDF, JPG, JPEG, PNG, TIFF, TIF, DOC, DOCX, PPT, PPTX, MD, and TXT files, and skips other types. File analysis runs in Agent mode, which uses multi-step reasoning to retrieve and analyze the relevant files. A workspace admin turns on the Beta feature Analyze Files in Volumes with Genie Agents from the Previews page.

Each Genie Agent supports up to 10 volumes, and each workspace can attach up to 250 volumes to Genie Agents. An indexed volume can contain up to 10,000 files, and Genie skips files over 100 MB or with more than 500 pages. Only Unity Catalog managed volumes are supported, not external or My Files volumes.

You can upload up to 25 files per conversation, regardless of type. Each CSV or Excel file must be smaller than 200 MB and contain fewer than 100 columns. Excel files can contain multiple sheets, and you name the sheet when you ask about it.

Only the user who uploaded a file can access it. Uploaded files have no fixed retention period and persist until you remove them, the conversation is deleted, or the agent is deleted. Users automatically receive USE CATALOG, USE SCHEMA, READ VOLUME, and WRITE VOLUME privileges on their own uploaded files.

PDF uploads require Agent mode. Each PDF must be under 20 MB and no more than 100 pages, with content limited to 15,000 characters. Uploaded PDFs apply to the current conversation and are uploaded through the UI, and partner-powered AI features must be enabled.

The external-tools page lists these MCP connectors for Genie Agents as Beta: Google Drive, Gmail, Google Calendar, Microsoft 365, Atlassian, Glean (Beta), Slack, and GitHub. The 2026 release notes list Databricks-provided MCP connectors for Genie One and Genie Code as generally available as of September 10, 2026, so check the current status for your surface. A workspace admin turns on MCP tools in Genie Agents from the Previews page, and each user authenticates to each connected system individually.

Every tool call runs with the permissions of the user who asks the question. Read-only tools such as search, fetch, and list run automatically, while write tools pause and ask for approval with the options Allow once, Always allow in this chat, or Reject. MCP tools run in Agent mode, which keeps the user in control of every write action.

Genie Code works in Databricks notebooks, the SQL editor, jobs, AI/BI dashboards, and the file editor. It answers questions with documentation citations, offers inline suggestions, and provides Quick Fix and Diagnose Error for failing code. Databricks recommends always reviewing generated code before running it.

Genie Code uses pay-as-you-go pricing with a per-user free monthly allowance, and charges apply beyond that allowance starting July 8, 2026 at 00:00 UTC. Service principals receive no free usage and are billed for all of their usage. Budgets apply to pay-as-you-go usage and can alert on or block spend, but they cannot remove a user's free monthly usage.

Only account admins can create and manage budgets. Create the budget with the Unity Gateway resource type and add the resource tag databricks-product: genie. Do not add other resource tags to a Genie budget, because additional tags prevent the budget from tracking Genie usage.

A budget supports up to 4 shared thresholds and up to 20 per-user overrides, and an account can have up to 1,000 budgets. Block usage is enforced approximately, and in-flight requests are not interrupted, so budgets help you manage spend. Use per-user thresholds to give each user a monthly limit, and overrides to raise the limit for specific users or groups.

Genie One is the simplified interface for business users, formerly called Databricks One, and it is a single entry point for data and AI. Chat searches across existing dashboards, queries, and Genie Agents to answer questions. Users interact with the data assets shared with them, and users with browse access can request access to a table from the table view.

Account-level Genie One does not aggregate data from workspaces with the compliance security profile enabled, so those workspaces stay separate. Some Databricks-generated metadata may be processed in the US. Account admins can disable account-level Genie One in the account console if that conflicts with their policies on US-based metadata processing.

Published dashboards include a Genie Agent by default, built from the dashboard's datasets and visualizations, when the Enable Genie toggle is on. To change this, open Settings > General on the dashboard and turn the toggle off, or select Link existing Genie Agent and paste its URL. Republishing the dashboard updates the autogenerated agent automatically to reflect new datasets and visualizations.

Dashboard filters do not carry over into the chat context of a linked Genie Agent, and Genie answers from the full dataset query results rather than only the visualized data. Define datasets carefully when they include sensitive columns or rows, so the agent sees only what you intend.

Companion Genie Agents support up to 30 dashboard datasets, and any datasets beyond 30 are not included. You cannot edit instructions for autogenerated agents in the Genie UI. Ask Genie is not available in external embedding, so use the Genie Conversation API for external applications.

Databricks now documents Genie spaces as Genie Agents. In the sidebar, select Genie Agents, click New, choose your data sources, and click Create. The data must be registered in Unity Catalog, the agent can include up to 50 tables, views, or metric views, and it requires a pro or serverless SQL warehouse on which you have at least CAN USE.

Yes. Example queries show Genie how to use the available data to answer questions, and when Genie receives a matching question it can use the example query directly. For similar questions it uses clues from the example to structure the SQL it returns. Add examples under Configure > Examples with a natural-language question and the SQL that answers it, and optionally add usage guidance explaining when the example applies.

Benchmarks are a set of test questions you run to assess Genie's overall response accuracy. In Chat mode, only questions that include a SQL Answer, the reference SQL statement, can be assessed automatically, and questions without one need manual review. After a run, Genie Code can review the results and suggest instruction and context improvements for you to review and save.

Common questions are optional example questions you add under Configure, then Settings in a Genie Agent. They appear on the Genie chat landing page, with your questions shown first. If you supply fewer questions than fit on the page, Genie fills the remaining slots with auto-generated questions. You can add them when you create the agent or later from Settings, which helps business users discover what the agent can answer.

Genie works best with specific, data-driven questions that it can turn into a single SQL statement. The documentation gives 'Why were sales low this quarter?' as a question Genie cannot handle well, because it does not interpret results or give recommendations. Genie may ask follow-up questions or say it cannot answer when context is missing, and responses can vary between runs for the same prompt.

Databricks recommends that most Genie Agents include between two and four phrasings of the same benchmark question, so accuracy is tested against natural variation in how people ask. Add a SQL answer wherever possible, because only questions with a SQL answer can be scored automatically. Questions without one need manual review in Chat mode.

No. Benchmark questions run as new conversations, so they do not carry the context of a threaded Genie conversation. Each benchmark result therefore reflects how the Genie Agent answers that question on its own, without influence from earlier turns in a thread. That keeps each question's result independent of how a conversation happened to unfold.

In Chat mode, Genie's generated SQL and result set are compared with the SQL answer you supply, not with the wording of its reply. Review each benchmark result and mark it Good or Bad, and see the benchmark documentation for the exact matching rules.

Chat mode compares up to 5,000 rows of each result set. When a result set exceeds 5,000 rows and the row order differs, the comparison can mark a correct answer as Bad. Treat a Bad rating on a large result as a prompt to check the mismatch manually before changing the Genie Agent's configuration.

Users with at least CAN EDIT permission on a Genie Agent can run all benchmark questions or a selected subset, and a run keeps going if you leave the page. The Evaluations tab lists timestamped runs with accuracy and the user who ran each one. Per-question details remain visible in the evaluation details for one week.

Agent mode does not compare SQL. An LLM judge grades each response instead. You can add an optional Evaluation note that describes the correct answer or expected content, and the judge uses it as grading guidance. Chat mode, by contrast, compares the generated SQL and result set with the SQL answer you supply.

Yes. After a benchmark run completes, you can use Genie Code to review the results across the whole run. It can then suggest context improvements for the Genie Agent. This turns a list of failed benchmark questions into a concrete set of configuration changes to test in the next run.

Users with CAN MANAGE permission can open conversations that are set to Reviewable by agent managers from the Monitor tab. For conversations a user has set to Private, managers can see only that user's prompts in the monitoring tab, not the full conversation or results. The Monitor tab also lets managers filter questions and answers by time, rating, user, or status.

Use the Weekly digest section of the Monitor tab to review weekly message volume and feedback. To find top usage trends and common issues, click Analyze Agent Usage, which uses Genie Code to report on them. This helps authors decide which instructions, example SQL, or data assets to improve next.

Shared conversations are live references, so recipients see the conversation as it currently stands. Recipients can forward the link, but they cannot change the conversation's sharing settings. This lets teams pass a Genie answer to colleagues without giving them control over who else can see it.

Yes. Click Download PDF on a completed Agent mode report to save it for sharing or offline review. The PDF includes the findings, visualizations, and citations to research steps. Agent mode uses multi-step reasoning for complex, exploratory questions, which makes the PDF a self-contained research output.

While Agent mode is reasoning, click Answer now to stop the reasoning and get an answer based on the context gathered so far. After a report is generated, you can also send additional clarification prompts to refine it. Use Answer now when speed matters more than depth.

Yes, if you add the allow=clipboard-write attribute to the iframe element. That attribute enables copying CSV data and conversation links. Get the iframe code from Share, then Embed space, after a workspace admin has configured the allowed embedding surfaces in Genie access management settings.

The visitor is prompted to authenticate before interacting with the embedded agent. Embedded users can send prompts but cannot change the agent configuration, and they see only the data they have permission to access. Agent authors need at least CAN MANAGE permission on the Genie Agent to set it up.

No on both counts. The server invokes Genie as a tool and does not pass conversation history to the Genie API. The Genie Agent MCP server is read-only: it queries data and does not write to your tables. Plan multi-turn flows in the calling application accordingly.

A workspace admin must enable Managed MCP Servers on the workspace Previews page. Then use https://<workspace-hostname>/api/2.0/mcp/genie/{genie_space_id} with the OAuth scope genie. Copy the space ID from the Genie Agent's Settings tab, and share the agent with the users or service principals that connect.

No. MCP tools run only in Agent mode and not in Chat mode. The external-tools page labels the connectors Beta for Genie Agents, while the 2026 release notes list the Databricks-provided connectors as GA for Genie One and Genie Code, so check the current status for your surface. A workspace admin turns on MCP tools from the Previews page, and each user signs in to each connector and acts with their own permissions.

The Google Drive connector supports only native Google Workspace files, meaning Docs, Sheets, and Slides, with a maximum file size of 25 MB. Other file types are not supported through this connector, so keep shared reference material in native Google formats when agents need to use it.

Not by default. The GitHub connector can reach only public repositories out of the box. Private repository access requires additional setup by a GitHub organization admin, and GitHub Enterprise Server is not supported by the connector. Plan to have that admin complete the setup before rolling out an agent that depends on private code.

A metastore admin must first configure the Glean tenant ID. Until that is done, non-admin users cannot use the Glean connection. Once it is configured, each user connects with their own account and tool calls run with that person's permissions, so access follows the user rather than a shared credential.

A Genie Agent always uses the permissions of the user who asks the question. Each user needs at least READ VOLUME on an attached volume, or they receive an error, and every user needs both the Workspace access and Databricks SQL access entitlements.

Databricks advises aiming for five or fewer tables and including only the tables needed for the questions the agent should handle, while limiting the columns in each. The hard ceiling is 50 tables, views, or metric views. When a topic needs more, pre-join related tables into views or metric views before adding them.

Add explicit instructions that state the source timezone, the conversion function, and the target timezone. Genie might not always be able to infer the timezone represented in the data. The troubleshooting guide's example converts UTC timestamps to Los Angeles time with convert_timezone.

Some regions require cross-Geo processing. If your workspace is in one of those regions, your account administrator must enable cross-Geo processing, according to the troubleshooting guide. This is an account-level setting rather than something an agent author can change, so ask your account administrator to enable it.

Call GET on the message attachment's query-result endpoint, using the space ID, conversation ID, message ID, and attachment ID. The endpoint is available only when the message has a query attachment and its status is EXECUTING_QUERY or COMPLETED. Read the answer text and generated SQL from the message attachments first.

Yes. POST to /api/2.0/genie/agents/{agent_id}/conversations/{conversation_id}/responses/{response_id}/cancel to cancel an Agent mode response. POST to /api/2.0/genie/spaces/{space_id}/conversations/{conversation_id}/messages/{message_id}/feedback to send feedback on a message, which lets applications capture ratings programmatically. Both calls let an application stop long-running Agent mode work and feed user ratings back into its own monitoring, using the same Conversation API as the rest of the integration.

Yes. The Genie app for Slack is documented as Public Preview. An account admin enables the Databricks Genie for Slack preview in the account console, and a Slack workspace admin installs the app. Users can message the app directly, mention @Genie in channels, or have a channel owner pin a specific Genie Agent for consistent answers. Genie One can rewrite a question before it reaches a pinned agent.

Responses are public by default, and setting message visibility to private limits the answer to the person who asked. Workspace-level settings override channel and personal settings. Account and workspace admins can monitor Slack usage in system.access.audit, where Slack requests appear as createGenieChatResponse events with request_params.source set to slack.

Yes. The Genie app in Microsoft Teams is documented as Public Preview. An account admin enables the Databricks Genie for Microsoft Teams preview, and a Microsoft 365 admin installs the app from the Microsoft marketplace and approves its permissions. Users can chat with it directly, @mention it in channels, or use it from Copilot. Channel owners run config to set the workspace, message visibility, and Genie scope.

The Teams documentation lists the points to plan for. You cannot choose which SQL warehouse the app uses, and the app does not connect to your workspace over Private Link. Triggering actions in Databricks from Teams is not supported. If your workspace uses an IP access list, allow the control plane's outbound IPs for the East US 2 region. Planning these up front makes the rollout smoother.

Depending on the source table, comments might not be accessible from Databricks, and Databricks does not manage the metadata, data, or semantics for writes to foreign tables. The documented fix is to edit the column metadata directly in the Genie Agent UI. That lets you set accurate business descriptions without changing the source system.

When the token limit is exceeded, you can no longer send or receive messages in that Genie Agent. The troubleshooting guide notes that unnecessary columns can significantly contribute to token usage, so trimming columns that no question needs helps keep the agent within its limit.

The Genie Agents troubleshooting guide gives two steps. Start a new chat when Genie ignores instructions or starts generating slow or failing responses. If responses are slow, check query history to identify slow-running queries, so you can see which statements take the longest and address them.

Yes. The 2026 release notes list Genie One memory as generally available as of September 29, 2026. Memory confirmation prompts were listed as Beta on September 10, 2026. Check the release notes for the current status of each Genie One enhancement before you plan a rollout.

Yes. The 2026 release notes list Databricks-provided MCP connectors for Genie One and Genie Code as generally available as of September 10, 2026. The release notes also list the GitHub MCP connection for Genie One as Beta as of July 9, 2026, so check the release notes for the current status of each connector before you roll one out.

Yes. The 2026 release notes say Genie One can now run code in a secure, isolated sandbox, listed as Public Preview on October 9, 2026. Confirm the current status in the release notes before you rely on the feature for production users.

Yes, in Beta. Genie Code CLI is a terminal-based agent that runs on your local machine and is preconfigured for Databricks data and AI work. Developers who prefer command-line workflows can use it for Databricks projects, and the Genie Code CLI documentation explains setup and usage.

Yes, in Beta. Genie One can search the public web for questions that need current information and shows each external source it uses as a link in the response. A workspace admin must turn on the Web search in Genie Code and Genie One preview from the Previews page first.

Yes. In Public Preview, when a Genie response includes a visualization, the chart appears as an image in the Slack reply. If charts are missing from public messages, a Slack workspace admin must reinstall the app to grant the files:write permission. Visualizations do not appear in private messages, which are visible only to the user who asked.

Yes. Genie Code scheduled tasks run a prompt on a recurring schedule. Each task executes the prompt as a full Genie Code session and produces a new chat that you can open and continue. The scheduled tasks documentation explains how to set one up.

Yes. Genie Code lets you fork a conversation into a new, independent thread from a point you choose. Use /btw to ask a side question without adding it to the thread history, and edit an earlier message to rewind the thread, which deletes the messages after it and regenerates the response from that point.

The full-page Genie Code experience is a command-center layout in which the active chat is shown prominently. Notebooks and files open as tabs beside the conversation, you can run multiple chats in parallel, and you can personalize Genie Code with skills, instructions, and MCP servers.

Add a task, set its type to Genie Code, and enter a natural-language prompt. The prompt can reference job parameters and upstream task values, and each run produces a continuable Genie Code conversation linked from the run output. Downstream tasks can read the final response from the task output. Auto-approval is always on for job tasks, so the agent runs without a manual approval step.

Genie Code CLI is a terminal coding agent that runs on your machine and is preconfigured for data and AI work on Databricks, including pipelines, models, and apps. It uses the Databricks CLI, version 1.0.0 or later, for authentication and genie commands such as genie ask. Model access comes through Unity Gateway, so you do not need your own model subscription. It is in Beta.

Yes, both renames are documented. The Genie One documentation states that Genie One was previously known as Databricks One, and describes it as the simplified Databricks UI for business users. The bundles documentation states that Declarative Automation Bundles is formerly known as Databricks Asset Bundles. When a page uses an older name, search the docs for the feature's current title and check the release notes for what changed.

Genie answers are generated by an LLM, so no text-to-SQL system is perfect, and Databricks recommends measuring accuracy rather than assuming it. Genie Agents benchmarks let you write test questions with ground-truth SQL and run them against the space: Chat mode compares Genie's query results to your supplied SQL answer, and Agent mode uses an LLM judge. When a benchmark fails, you inspect the generated SQL, fix the instructions, example queries, or metadata, and rerun the suite before the space is used for important numbers. Passing benchmarks shows accuracy on the questions you tested, so keep human review for high-stakes answers.

Databricks documents that Genie, like other large language models, can exhibit non-deterministic behavior, so the same prompt can produce different outputs. The docs recommend that most Genie Agents include two to four phrasings of the same important question, so you test whether the answer holds across wordings. Clear text instructions, example queries, and SQL expressions for business logic reduce the variation, and the tuning guidance says defining business logic as SQL expressions produces more consistent results than text instructions.

Databricks states that a Genie Agent's behavior does not change based on user feedback alone, so feedback is input for the people who maintain the space rather than automatic retraining. Users can flag a response as incorrect with the Fix it option, and the fix has to be made in the space's instructions, example SQL, or metadata. Genie can also suggest new SQL expressions after positive feedback. Once a fix is in, rerun your benchmark questions to confirm the change actually improved accuracy.

For repeatable checks, Genie Agents benchmarks compare results against a validated ground-truth SQL statement in Chat mode, and use an LLM judge in Agent mode. In Chat mode, only questions that include an example SQL statement can be assessed automatically for accuracy, so write and validate SQL for your most important metrics first. Benchmark questions and SQL are used only for evaluation and do not change Genie's context, so add any useful corrections separately as instructions or example queries. Treat benchmark results as evidence about the questions you tested, not a guarantee for every question a user might ask.

Databricks recommends putting global business context in text instructions, which apply to every prompt, and putting question-specific logic in example queries, SQL functions, or metadata. For recurring KPIs, the Tune Genie Agent quality guidance says defining the logic as SQL expressions produces more consistent results than text instructions. Parameterized example SQL gives verified answers in Chat mode when a user's question uses the exact text of the query. Pair these with clear table and column descriptions in Unity Catalog, and keep instructions non-conflicting so Genie has one definition to follow.

A trusted asset is a parameterized example SQL query or Unity Catalog SQL function whose exact logic has been verified by an agent author, so when Genie uses one, the answer comes from that verified logic. It is not a blanket guarantee: Databricks positions benchmarks as the way to measure Genie Agent response accuracy, and the concepts page makes no guarantee that Genie is correct. Test the phrasings your users actually use, and start with the handful of metrics your business asks about most often.

Genie demos often fail on real questions because the space lacks business context. Databricks' docs recommend tuning table and column descriptions, join relationships, example SQL, SQL expressions, and general instructions. Measure accuracy with a benchmark set of test questions with SQL answers (up to 500 per agent), and iterate from the evaluation results.

Genie works best on curated, well-modeled data. The docs say Genie operates in a nondeterministic manner and recommend iterating on the agent based on testing and usage. Plan metadata, relationships, example SQL, and instructions, and use a benchmark set to measure accuracy before business users rely on it.

Genie Agents benchmarks are a set of test questions you define, with optional ground-truth SQL, used to evaluate and measure response accuracy. Because the questions and expected SQL are fixed, you can rerun the same suite after each change to instructions, example queries, or metadata and compare the results. Benchmark questions and SQL are used only for evaluation and do not change Genie's context, so new knowledge has to go in as instructions or example queries. Keeping the suite under version control with your space changes makes regressions visible before users see them.