Library limits for sizing files, batches, and storage plans

Use this page to size a file, a corpus, a batch, or a storage plan. It groups the Library limits that affect those decisions by the question that sends you looking for them.

Some limits are part of the versioned API contract. Changing them breaks clients. Other limits are service settings that a release can change.

What one file can be

One file can be 500 MiB at most, and its file type must be one the platform reads. Split a larger source and upload each part as its own document.

The platform reads 143 file extensions in 9 categories. These counts were verified on 2026-08-31:

Two notes on what that list covers, and two boundaries:

• Documents include .pdf, .docx, .doc, .rtf, .odt, .epub, .html, .md, .txt, .tex, .pages, .xps, .cbz, .mhtml, and 11 more extensions.
• Data covers .xml, .json, .jsonl, and .ndjson. Code covers 56 extensions.
• From a video the platform reads the text that is visible in the frames. It does not read the speech.
• The platform reads text, so a recording enters a Library as a transcript or as a subtitle file. Subtitles cover .srt, .vtt, and seven more extensions.

GET {DUALE_ENDPOINT}/libraries/v1/tenants/{tenant_id}/document-uploads/supported-formats returns the exact advertised extension list and is the authority for a file picker. The route checks the library:upload action, its response is cacheable, and the list changes only with a release. Read the route before a large import; the counts above are orientation for the date they name.

What a Library can hold

Two values bound a Library: its path length and its tags size. Nothing caps the documents in a Library or the Libraries in a tenant, so size a Library around the agent that reads it, not around a count.

What one call returns

Every listing and every document tool returns a bounded result.

The model’s four document tools run under ceilings that your code does not pass:

Your own code passes none of those numbers. The model chooses a tool during a task, and the platform applies the ceiling. What agents can do with a Library explains what each tool matches, why a trace name is not an API, and how to read an answer that stopped early.

Two numbers belong to your own code instead:

• list_documents(request) takes a limit of 1 to 500 and returns that many documents in one response, with a cursor for the next response. Pass each cursor back until next_cursor is None.
• list() returns the Libraries the caller can read. When the caller can reach 5,000 resources, the call raises an error instead of returning a partial list. The platform exposes no counter for that set. Treat the error as the signal: split the caller’s work across agent identities with narrower Library grants, then retry.

SDK reference states both signatures and the exception type, and Access and isolation states who grants the actions that a caller resolves.

How long each window lasts

Four windows govern incomplete uploads and deletion. Act before each one ends.

Nothing removes a document permanently inside its 30-day window unless you delete its Library. Deleting that Library also deletes the document, even while it is in Recently deleted. Security and data handling covers what the platform does with a document between the delete and the removal.

How long processing takes

The SDK’s polling helper stops after 360 seconds by default, before the platform’s processing allowance can end. A client timeout does not prove that processing failed. Pass a longer timeout= for a large file, then read the returned status instead of uploading the file a second time.

The platform allows 30 minutes to extract one document and 15 minutes to make each part of its text searchable. A longer document has more parts, so do not compute a total wait from those two numbers. Manage libraries shows the polling call and the states a document moves through.

Duale AI publishes no total time for a batch, and documents can finish in any order. Test a representative batch under your production conditions, then wait for every document the task needs to reach ready.

What the platform meters

The platform meters stored document volume over time.

A document counts from when the platform stores it until the platform deletes it permanently. A deleted document stays in the 30-day Recently deleted window and still counts during that window. Delete it as soon as your retention policy allows. Include the 30 days in its cost.

Nothing expires a Library that a task created for its attachments, so its documents count until you delete them. Attach documents covers the files that one task carries. Pricing and access states how the order form fixes the storage rate and included volume.

What a nightly refresh costs

A nightly refresh that uploads every file and deletes the prior documents still holds about thirty superseded copies of every file, plus the current one. A change-only refresh keeps that extra volume to the files that changed.

This happens because uploading a file again creates a new document instead of replacing the old one, and a document you delete keeps counting for its 30-day window.

Best practices gives the habit that keeps a changing set of documents to roughly its own size.

What Libraries do not do

Libraries have eight current boundaries. Draw your design around them.

Read-back is the boundary behind every diagnosis. You cannot replay the passages an agent read, so you test a corpus with a question you know one document answers, then read the answer. SDK reference states the preview call, and Best practices gives one check for each habit.