Databricks gives Genie Agents an indexed path through unstructured files

Databricks has added optional content search for Unity Catalog volumes attached to Genie Agents, moving persistent document analysis away from purely query-time parsing and toward a pre-indexed retrieval path. The capability entered Beta in the company’s August 13 release notes, alongside removal of the requirement to enable Unity AI Gateway for analyzing files in volumes. Databricks release notes

Two retrieval paths

A Genie Agent can attach as many as 10 Unity Catalog volumes and answer questions across their unstructured contents and structured tables in one conversation. File analysis runs only in Agent mode: the agent retrieves relevant content, combines it with structured data, and returns an answer with citations. Access is evaluated with the querying user’s permissions, including a requirement for at least READ VOLUME on the attached data. Databricks documentation

Without content search, the constraints are relatively tight. The agent can retrieve context from no more than five files per question; a volume can contain up to 500 counted files, and supported files are limited to 10 MB. Databricks says questions against larger non-indexed volumes return an error. Databricks documentation

Enabling content search changes the operating envelope. Databricks says the service prepares and indexes files to improve latency and reasoning over larger collections. An indexed volume can contain up to 10,000 files, each as large as 50 MB. The documentation recommends indexing attached volumes, while still describing the option as part of a Beta feature that an administrator must enable. Databricks documentation

The cost model changes too

The choice is not only about scale. During the Beta, a non-indexed volume is billed for the agent’s reasoning over content parsed at query time. With content search enabled, the agent queries an existing index, so Databricks says retrieval does not incur Foundation Model Serving costs; ingestion and query-serving charges apply instead. Standard Genie reasoning charges remain in both cases. Databricks documentation

That makes content search a deployment decision rather than a cosmetic toggle. Small or temporary collections can stay on the query-time path, while larger persistent corpora can be indexed in advance. The limits remain material: the feature is unavailable on AWS GovCloud, cannot analyze external sources such as Google Drive or SharePoint, and requires both Workspace access and Databricks SQL access entitlements for every user. Databricks documentation

For NL2SQL systems, the release broadens the context surrounding a generated query. Genie can now retrieve supporting policy files, call notes or support documents alongside governed tables, but the quality of that context still depends on volume descriptions, file naming and removal of duplicate or irrelevant documents—the curation steps Databricks explicitly recommends. Databricks documentation