Short answer: since November 2, 2026 the Databricks AWS release notes list file analysis in Unity Catalog volumes as Public Preview. A Genie Agent in Agent mode can answer from twelve filename extensions, including PDF, with citations that can point to a source page, using the asking user's permissions. Documented limits include 10,000 files per indexed volume and skipped files over 100 MB or 500 pages. The volumes guide states no models, retention, encryption, accuracy or GA date. We tested nothing.
Databricks is a data platform for storing and analyzing business data. Genie Agents are its AI assistants, and the documentation describes one mode of them, Agent mode, as multi-step reasoning. The new feature lets an assistant read documents such as PDFs, slide decks and images kept in a volume, which is a governed storage location for files, and answer with a reference to the page it used. A logistics or sales team might wonder whether it could help with carrier paperwork, contracts or manuals.
The catch: the feature is in Public Preview, it takes a whole volume rather than chosen folders, and the documentation sets caps on file counts, sizes and pages. The guide itself does not name the AI models, say how long data is kept, or give accuracy or price. This article lists what is stated and what to ask.
We read the Databricks documentation on November 4, 2026 and tested nothing. This teardown covers one feature: a Genie Agent answering questions from files in Unity Catalog volumes. Its main source is the volumes guide, "Analyze files in volumes with a Genie Agent", plus the AWS release notes. For billing, permissions, audit and data handling we also read Databricks pages on Genie pricing, volume privileges, audit logs, AI trust, partner-powered models and ai_parse_document; each statement is scoped to its own page and the pages are listed under sources. Our inference is marked as ours, and the questions about invoices, bills of lading and manuals are our questions, not Databricks claims.
What did Databricks release on November 2?
The AWS release notes have an entry dated November 2, 2026, titled "Analyze files in Unity Catalog volumes (Public Preview)". It says a Genie Agent can analyze unstructured files in volumes, such as PDFs, slide decks and images, and answer with citations that point to the source page, and that workspace admins control access from the Previews page. The entry carries no limits, regions, billing, model or accuracy text; the detail sits on the guide. The notes also warn: "Releases are staged. Your Databricks account might not be updated until a week or more after the initial release date." Our inference: November 2 does not guarantee the preview is visible in every eligible workspace that day.
On status: earlier entries for this feature, dated July 23 to October 29, are labelled Beta, so in the AWS notes the label moved to Public Preview on November 2. We did not open the Azure or GCP pages. The October 16 entry on file upload is a different feature, which that entry calls generally available.
What is a Unity Catalog volume, and how does the agent read files?
Our plain-words gloss: a volume is a governed storage location for files such as PDFs and images, kept in cloud object storage and controlled through Unity Catalog with its own privileges, such as READ VOLUME. Databricks' volume pages say volumes govern non-tabular data. According to the guide, the agent retrieves the file content most relevant to the question, adds it to its context, then reasons over it together with any structured data to produce an answer with citations. Users can ask across one volume, several volumes and structured data in the same conversation.
Three conditions shape the rest. File analysis runs only in Agent mode, which the docs describe as multi-step reasoning. The agent takes a whole volume: you cannot attach a subset of the folders or files in it. And when you attach a volume, Databricks automatically enables "content search", the step that parses and indexes your files so the agent can work across more of them with lower latency and keep in sync as files change. The guide also says analysis is available through the Agent mode APIs.
What are the documented limits?
Most numeric limits are phrased around content search, the indexing step, not the agent in general. The guide does not say what the agent does with a file over a cap, so we write "content search skips", not "the agent cannot read".
| Item | What the docs state |
|---|---|
| Volumes per agent | A maximum of 10 |
| Volumes per workspace | Up to 250 attached to Genie Agents |
| Files per volume | Each indexed volume can contain up to 10,000 files |
| Size and page caps | Content search skips files larger than 100 MB and files with more than 500 pages |
| Filename extensions | Twelve: PDF, JPG, JPEG, PNG, TIFF, TIF, DOC, DOCX, PPT, PPTX, MD and TXT; all other types are skipped |
| Volume type | Only Unity Catalog managed volumes; external volumes and My Files volumes are not supported |
| Subsets | None: the whole volume is processed |
| Mode | Agent mode only |
| Users | Every user needs both Workspace access and Databricks SQL access entitlements; during Public Preview, account-only users cannot use the feature |
| Platform | Not available on AWS GovCloud; Unity Catalog must be enabled in the workspace |
| Regional prerequisites | Agent mode, Lakebase, ai_parse_document and Foundation Model APIs must be available in your region |
| External sources | For content from sources such as Google Drive or SharePoint, the docs say to attach MCP connectors instead |
That is twelve filename extensions, not twelve formats: JPG and JPEG, and TIFF and TIF, are variants. Spreadsheets, audio and video are not on the list and are skipped.
Whose permissions does the agent use?
The guide says the Genie Agent always uses the permissions of the user who asks, and a user without at least READ VOLUME on an attached volume receives an error. Our inference: the people who can read a volume can read what the agent says about it, so review the reader list before attaching anything.
The guide names READ VOLUME as the minimum and mentions no other privilege. Databricks' general Unity Catalog documentation, not the Genie guide, lists USE CATALOG on the catalog, USE SCHEMA on the schema and READ VOLUME on the volume for reading files, and says privileges can be set at catalog or schema level and cascade to the volumes inside. We do not know whether the agent needs more than READ VOLUME, so include inherited grants in the reader list.
The guide has no audit section. Databricks documents Genie Agent audit events separately, logged under the service name aibiGenie. In the part of that reference we read, no event name mentions volumes or content search, so whether an administrator can see which files an agent used for an answer is not stated in what we read. For the same question about a different kind of tool, see our agent-desktop teardown.
Where does the processing run, and what does it cost?
For eight regions the guide says content search runs in your workspace region: Asia Pacific (Sydney), Asia Pacific (Tokyo), EU (Frankfurt), EU (Ireland), EU (London), US East (Northern Virginia), US East (Ohio) and US West (Oregon). For four, Asia Pacific (Mumbai), Asia Pacific (Singapore), Canada (Central) and South America (São Paulo), an account admin must enable cross-Geo processing before content search runs, which allows Databricks to process your files outside your workspace Geo. The guide says content search is available in the regions it lists, twelve in all, and names no other AWS region as supported. The guide does not say where the answer-generation step runs, so we do not claim files stay in a region.
On billing, the guide says that when Databricks enables content search or syncs a volume, it parses your files with the ai_parse_document AI function, billed under AI Functions pricing. The agent's reasoning is billed as part of standard Genie usage. The guide gives no price or worked example. It says that when you add, change or remove files, Databricks updates the index, but not which files are re-parsed or what a re-sync costs. That we do not know, and we will not guess a figure.
Separately, the AWS release notes of August 3, 2026 say free usage of Genie One and Genie Agents is extended through January 31, 2027, for usage by users only; service principals are excluded and still billed. Databricks' Genie pricing page footnotes that free usage applies to Genie LLM usage only and that Genie compute, for example SQL Serverless, is billed separately. The August 3 entry does not mention parsing or AI Functions, so we do not claim the free period covers parsing. Promotions change.
What the volumes guide does not say
We checked the guide by keyword and read its headings. It states none of the following. Silence is not a denial.
- Models. It does not say which models read or reason over the files; Foundation Model APIs appear only as a regional requirement.
- Training use, retention and encryption. Nothing on training use, how long parsed content or the index persists, where it is stored, or how it is encrypted. The index updates when files change or are removed; duration is not stated.
- Accuracy and latency. No accuracy figures or benchmarks, no latency numbers beyond "lower latency", and nothing on reading scans, photographs or handwriting.
- General availability. No date or roadmap, only Public Preview.
- Price and re-sync cost. No numbers.
What other Databricks pages say
Separate pages about AI features in general cover some of these topics. None mentions volume content search, and whether this feature falls under their statements is not stated, so treat them as leads for your questions.
- Models. The partner-powered page lists "OpenAI on Databricks and Anthropic on Databricks" for Genie Agents in Agent mode when the partner-powered AI features setting is enabled (the table says Agent mode is not available when it is disabled), with Anthropic on Databricks using endpoints hosted by Databricks within the Databricks security perimeter. It does not mention file analysis, so it does not say which model reads your files.
- Training, partner retention and encryption. The AI trust page says Databricks does not use data, prompts or responses submitted to or output from "these features" to train generative foundation models it makes available to third parties; that model partners do not retain submitted data, even for abuse monitoring; that traffic to model partners is encrypted in transit with TLS; and that data stored within a workspace is AES-256 bit encrypted. Its list of Genie Agent data does not mention files or volumes.
- Parsing. The
ai_parse_documentpage says your document data is processed within the Databricks security perimeter and that parameters passed into the function are not stored, though metadata run details such as the runtime version are retained. It does not describe the content search index.
Still unknown after reading all of these: how long the index is retained and how it is deleted, the general availability date, and whether a file-level audit record exists.
Do this today
Three steps that take an afternoon and need no developer.
- Pull a sample of your worst scanned or fax-quality documents and note which are PDFs, images or spreadsheets.
- Count your files and the longest page counts, and compare them with the stated caps for files per volume, size and pages.
- Write the vendor questions below in an email and ask for answers in writing before any pilot.
Questions to put to Databricks or your data team
These are ours, for logistics, sales, retail and operations teams. The guide's own examples are support escalations, stalled sales opportunities and underperforming regions. It names no invoice, bill of lading or equipment manual, so treat those as something to test, not something the docs promise.
- Our shipping documents sit in folders under one share. The agent takes a whole volume. Can we restructure storage so a volume holds only what an agent should see?
- We have more than 10,000 invoices and bills of lading a year. With 10,000 files per indexed volume, 10 volumes per agent and 250 per workspace, how many volumes would we need, and would a split by year fit?
- Many carrier PDFs are scans or fax quality, and the guide says nothing about scan quality. Who tests a sample of our worst documents before a pilot, and what error rate on totals and container numbers is acceptable?
- Some equipment manuals may exceed 500 pages or 100 MB. Which fall outside the content search caps, and how would we split them?
- Our files live in external storage, SharePoint or spreadsheets. Only managed volumes are supported, spreadsheets are not listed and external sources need MCP connectors. Managed volumes sit in cloud storage governed by Unity Catalog, so must our files move into one, and is that acceptable for compliance and cost?
- Who audits who holds
READ VOLUMEand the inherited catalog and schema grants, and can we see which files an agent used for an answer? - If our workspace is in Mumbai, Singapore, Canada (Central) or São Paulo, does our contract allow files to be processed outside the workspace Geo, and who signs that off? If our region is not among the twelve listed, is content search available at all?
- Before putting customer invoices in a volume, what do we ask Databricks in writing about models, index retention and deletion, and whether the trust page statements cover volume files? Which files are re-parsed on a re-sync, and what does it cost?
- It is Public Preview with no stated GA date. What is our fallback if terms change?
Citations that can point to the source page make spot checks possible; the guide does not say how reliable the answers are, so the sampling is ours to design. For wider context see our comparison of enterprise AI agent platforms and our piece on production AI agents architecture.
What is still unknown?
The guide shows "Last updated on Oct 30, 2026", two days before the November 2 entry that links to it; we do not know why, so we do not say when it last changed. We did not open the Previews, cross-Geo, Designated Services or Agent mode API pages, and we use no figures from any pricing table. We read only the first part of the long release-notes and audit-log pages. We did not run the feature and have no data on adoption or results. This article reflects what was public on November 4, 2026.
Sources
- Databricks documentation: Analyze files in volumes with a Genie Agent (AWS), read November 4, 2026
- Databricks AI/BI release notes 2026 (AWS), November 2 and August 3 entries and staged-release note, read November 4, 2026
- Databricks: Genie pricing, footnote on free usage, read November 4, 2026
- Databricks documentation: What are Unity Catalog volumes?, read November 4, 2026
- Databricks documentation: Unity Catalog volume privileges, read November 4, 2026
- Databricks documentation: audit log reference, Genie Agent events (first part read), read November 4, 2026
- Databricks documentation: Databricks AI trust and safety, read November 4, 2026
- Databricks documentation: partner-powered AI features, read November 4, 2026
- Databricks documentation: ai_parse_document function, read November 4, 2026
- Databricks blog: Expanding Genie Agents, background only, read November 4, 2026; no limits or status taken from it