Databricks Genie Agents now answer from PDFs and slides in Unity Catalog volumes

Databricks has added a Public Preview that lets a Genie Agent read unstructured files stored in Unity Catalog volumes, such as PDFs, Word documents, slide decks and images, and answer questions with citations that point to the source page. The change appears in the AI/BI and Genie One release notes under November 2, 2026, and the documentation describes it as a way to combine file content with the structured tables an agent already uses in a single conversation.

For a BI team, the practical effect is that a Genie Agent built over sales or support tables can also draw on the policy documents, call summaries or product manuals that explain those numbers, without a separate retrieval pipeline. The feature is off until a workspace admin enables "Analyze Files in Volumes with Genie Agents" on the Previews page, it requires Agent mode, and it only indexes Unity Catalog managed volumes, so external volumes and My Files volumes are excluded.

The preview arrives seventeen days after Databricks made conversation file upload in Genie One and Genie Agents generally available on October 16, 2026. The two features cover different needs: uploads serve one user and one conversation, while volumes give every authorized user of an agent the same governed document collection, which the agent indexes and keeps in sync as files change.

··········

HOW GENIE AGENTS USE FILES IN UNITY CATALOG VOLUMES.

An agent author attaches one or more volumes to an existing Genie Agent. According to the documentation, the steps are to open the agent, open the Sources tab, click Add, select the volumes and write a clear description in the Description field. The agent processes the entire volume; Databricks states that a subset of folders or files cannot be attached. A volume can be attached even when its catalog is not accessible from the workspace by default, for example a catalog bound to specific workspaces.

Attaching a volume automatically turns on content search, which parses and indexes the files so the agent can reason over more documents, pull specific passages from many files and respond with lower latency. Databricks says the index updates on its own when files are added, changed or removed. When a question needs document content, the agent retrieves the most relevant passages, reasons over them together with any structured data, and returns an answer with citations that can identify a specific page of the source file.

Users ask file questions from the chat box with Agent mode turned on, which the documentation lists as mandatory for this kind of question. One conversation can query a single volume, several volumes and the agent's tables together, so a question that compares a revenue trend with the terms of a pricing policy can be answered in one thread. The same capability is exposed through the Agent mode APIs for teams that embed Genie Agents in their own applications.

Databricks positions the feature for specialized, domain-focused agents. For a broader entry point it points to Genie One, which can call a Genie Agent for file questions, and for control over model selection, chunking strategy or several sub-agents it points to Multi-Agent Supervisor.

··········

REQUIREMENTS AND REGIONS FOR GENIE FILE ANALYSIS.

The documented prerequisites go beyond the preview switch. Unity Catalog must be enabled, Agent mode must be available in the workspace region, and Lakebase, the ai_parse_document function and Foundation Model APIs must also be available there. Every user of the agent needs the Workspace access and Databricks SQL access entitlements, plus at least READ VOLUME on each attached volume. During the preview, account-only users cannot use the feature, and it is not available on AWS GovCloud.

The AWS documentation page splits supported regions into two groups. In eight regions content search runs in the workspace region without cross-Geo processing, while four regions require an account admin to enable cross-Geo processing first.

........

Region group	Regions listed by Databricks
Content search in the workspace region	Sydney, Tokyo, Frankfurt, Ireland, London, N. Virginia, Ohio, Oregon
Cross-Geo processing required	Mumbai, Singapore, Canada (Central), São Paulo
Not available	AWS GovCloud

........

Teams with data residency obligations should treat the second group as a governance decision rather than a configuration detail, because enabling cross-Geo processing means document content may be processed outside the workspace's geography.

··········

FILE TYPES AND LIMITS FOR GENIE AGENT VOLUMES.

Supported formats are PDF, JPG, JPEG, PNG, TIFF, TIF, DOC, DOCX, PPT, PPTX, MD and TXT. Databricks states that all other types are skipped, which means CSV and Excel files sitting in the same volume are not indexed by content search; tabular files belong in tables or in conversation uploads. The documented ceilings are 10 volumes per Genie Agent, 250 volumes attached to Genie Agents per workspace and 10,000 files per indexed volume, and files larger than 100 MB or longer than 500 pages are skipped.

Because the documentation says unsupported, oversized or overlong files are skipped, an author should check a volume against these limits before relying on it. The illustrative inventory below applies the documented rules to seven files; the results were computed with a short Python script.

........

Illustrative file	Type	Size	Pages	Result under documented limits
policy_handbook_2026.pdf	PDF	4.2 MB	180	Indexed
q3_sales_review.pptx	PPTX	38 MB	46	Indexed
contract_scan_archive.pdf	PDF	140 MB	320	Skipped, over 100 MB
field_photos_0412.png	PNG	6.5 MB	1	Indexed
legacy_manual.pdf	PDF	22 MB	640	Skipped, over 500 pages
support_export.csv	CSV	12 MB	n/a	Skipped, type not supported
meeting_notes.md	MD	0.1 MB	n/a	Indexed

........

Four of the seven files would be indexed. The large scanned archive and the long manual would need to be split before upload, and the CSV export would be better loaded into a table the agent already queries. The same pre-check fits in a small Python function that can run over a volume's file listing.

SUPPORTED = {"PDF", "JPG", "JPEG", "PNG", "TIFF", "TIF", "DOC", "DOCX", "PPT", "PPTX", "MD", "TXT"}

def genie_volume_status(file_type, size_mb, pages=None):
    if file_type.upper() not in SUPPORTED:
        return "skipped: type not supported"
    if size_mb > 100:
        return "skipped: over 100 MB"
    if pages is not None and pages > 500:
        return "skipped: over 500 pages"
    return "indexed"

print(genie_volume_status("PDF", 140, 320))  # skipped: over 100 MB

The documentation's best practices point the same way: describe each volume's content and intended use, avoid overlapping content across volumes, keep only files relevant to the agent's domain and use descriptive file names.

··········

PERMISSIONS AND COSTS OF GENIE FILE ANALYSIS.

The agent always runs with the permissions of the user asking the question. A user without READ VOLUME on an attached volume receives an error, so access to documents follows Unity Catalog grants and does not depend on who built the agent. In practice, an agent shared with a broad sales audience should only be attached to volumes that the whole audience is allowed to read, or the experience will break for part of it.

Billing has two parts. Preparing files runs the ai_parse_document AI function, billed under AI Functions pricing whenever content search is enabled or a volume syncs, and answering questions is billed as standard Genie usage. Databricks does not publish a per-file figure on the feature page, so the cost of indexing depends on the volume's size and how often its files change.

A related platform change took effect on November 1, 2026: Unity Gateway budgets now include spend from task-specific AI Functions such as ai_parse_document, ai_extract and ai_classify, across batch and real-time execution paths. Admins who already track AI spend through those budgets will see document parsing for Genie volumes counted alongside other AI Functions workloads. Similar document parsing functions have reached other platforms recently, as covered in our report on BigQuery AI.PARSE_DOCUMENT becoming generally available and in the Snowflake AI_EXTRACT update for 1,000-page documents.

··········

VOLUMES VERSUS CONVERSATION UPLOADS IN GENIE.

The generally available upload feature and the new volume preview overlap on PDFs but differ in scope, limits and governance. The comparison below uses only figures stated on the two documentation pages.

........

Aspect	Conversation file upload (GA)	Files in volumes (Public Preview)
File types	CSV, Excel, PDF	PDF, images, DOC, DOCX, PPT, PPTX, MD, TXT
Scope	One user, one conversation	All users with READ VOLUME
Quantity	Up to 25 files per conversation	Up to 10,000 files per volume
PDF limits	Under 20 MB, up to 100 pages	Up to 100 MB, up to 500 pages
Indexing	None, used in the conversation	Content search, kept in sync
API	PDF uploads are UI-only	Agent mode APIs

........

Uploads remain the right tool for a quick look at a spreadsheet a user has on a laptop, while volumes suit reference material that many people should query in the same governed way. On the structured side, Databricks has also been extending Unity Catalog modeling, for example with the relationships between metric views now in Beta.

··········

DATABRICKS GENIE VOLUME ANALYSIS: WHAT TO CHECK BEFORE ROLLOUT.

Databricks confirms the core facts: Public Preview status, admin enablement on the Previews page, Agent mode as a requirement, managed volumes only, page-level citations, the listed file types, the per-agent, per-workspace and per-volume ceilings, and billing split between AI Functions pricing for parsing and standard Genie usage for answers. Availability depends on the workspace region, on whether an account admin accepts cross-Geo processing in four regions, and on regional availability of Lakebase, ai_parse_document and Foundation Model APIs.

What the documentation does not state is a general availability date, a per-file or per-page indexing cost, or the delay between a file change and an updated index. For external sources such as Google Drive or SharePoint, Databricks directs teams to MCP connectors instead of volumes. Teams comparing assistant-led analysis across platforms can set this preview against the approach described in our coverage of the OpenAI Data agent in ChatGPT Work for BigQuery, Snowflake and Databricks, keeping in mind that Genie's document access is bounded by Unity Catalog grants on each volume.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org