Oracle adds file uploads and downloadable outputs to AI agent conversations

Oracle is adding a way for business users to bring a fresh report into an AI agent conversation and leave with a downloadable result. Announced November 10, support for reading and generating files starts in November 2026 in Oracle AI Data Platform, with developers controlling file handling, storage and access.

A fresh report becomes conversation context

The feature addresses information that arrives after an agent has been configured: a sales export, supplier report, invoice or screenshot. Oracle describes using those files as additional context even when their contents have not been added to the data systems the agent already uses. The user supplies the new material directly in the conversation.

Oracle illustrates the workflow with a sales manager uploading a PDF report and requesting a five-bullet summary. The manager can ask which actions the report mentions, shorten the response, then request a text file containing the revised summary. This is a company walkthrough, not a measured customer result. It shows a sequence in which the user reviews and refines the content before requesting an export.

Developers decide what the agent can ingest

File Read and File Write are enabled independently through the Chat Trigger, the entry point for a user's message. Oracle's documentation makes a key constraint explicit: supported file types depend on the selected model. Developers must choose a model that supports their intended inputs and outputs; enabling the settings does not establish universal format support.

For incoming files, developers select accepted formats, a destination folder or volume, maximum size and the number allowed in one request. They can also offer an upload preview so users can check the document before processing. Oracle's example accepts PDF, TXT and CSV files, with a 5 MB size setting and a five-file allowance. Those are walkthrough choices, not stated platform-wide limits.

Processing controls determine how the file is prepared. Data enrichment extracts useful text and metadata, including titles, sections and document structure. Truncation settings determine which portion is retained when a file exceeds the relevant processing limit. These settings govern the material the agent receives, rather than simply whether an attachment is accepted.

Keeping a file and sharing it are separate decisions

Developers can choose whether uploaded files remain available across sessions. With persistence disabled, the agent can retrieve their information only during the current session. With persistence enabled, developers specify how long files are retained. Reusing a report in a later conversation is therefore a configuration decision, not an automatic consequence of uploading it.

For generated files, developers set output formats and maximum size, with optional storage in the AIDP Catalog at a selected folder or volume. Oracle's walkthrough permits PDF, TXT and CSV output with a 1 MB maximum. Again, that number describes the example configuration, not a platform-wide ceiling. A user can request a revised artifact, subject to the agent's configured capabilities and limits.

Output controls also cover retention, encryption, visibility and download-link expiry. Visibility determines whether access is limited to the end user or includes the developer. Link expiry controls how long a particular download link works. That is distinct from the persistence and retention settings governing whether the underlying file remains available for reuse and how long it is kept.