Libraries: document retrieval, access, and corpus design

A Library is a durable collection of documents in your Tenant. You upload a file once. During a Task, the model can retrieve material from the Libraries that the Task’s Agent Identity can reach. You write no retrieval code: no text splitting, no embedding job, no vector database.

What a Task can do with a Library

Four Library Tools are Built-in Tools:

• List what is there.
• Find the passage—the piece of a page—that means what the Task asked, across your working languages.
• Match literal text, such as an invoice number or an error code, ignoring case and accents.
• Read back the lines it needs to quote.

The model chooses among them, and your code calls none of them. An available Tool can go unused during a Task. Retrieval identifies material that can be relevant; it does not establish that the material is true or sufficient evidence for a claim.

Who can reach a Library

An allow Grant or a Tenant access policy can give an Agent Identity access to a Library, and a refusal on the Library overrides both. By default, every Agent Identity reaches the Libraries under its own agent/<agent_id>/ path, and supplied policies give people access by tier.

A Grant does not keep Library content away from people who hold supplied tiers: the Viewer and Editor tiers read every Library in the Tenant. Use a separate Tenant to keep documents away from a person. Access and isolation owns these rules.

Other names for this capability

Other platforms sell this capability under other names. A reader who arrives looking for document search, agent knowledge, a knowledge base, a vector store, or retrieval-augmented generation (RAG) is in the right section. Two differences are worth stating early, because they change how you design a corpus:

• The literal match is its own tool, not a setting inside search. An invoice number is found literally, not approximated, because the model asks for the text itself instead of a relevant passage.
• There is no query-time attribute filter. The Library is the scope. Put the documents that answer one kind of question in one Library, and grant that Library to the Agent Identity of the Tasks that need it.

Continue from your current Task

Understand retrieval

Library retrieval shows how each Tool matches, how to read an answer that stopped early, and how a document becomes searchable.

Design the corpus and access boundary

Design your Libraries helps you choose how many Libraries to create and which Agent Identity can reach each. Use Access and isolation to grant, refuse, and review access; Security and data handling to review the checks, deletion, and records; and Limits to look up sizes, ceilings, windows, and what Duale AI measures. Best practices helps you prepare documents that index well and keep answers complete.

Integrate Library operations

Manage libraries covers the client calls that create a Library and upload a document. Attach documents covers the files one Task carries.

Diagnose a failure

When a document will not load, SDK errors lists each failure code and its recovery. When a ready document gives a wrong or incomplete answer, start with Troubleshoot a missing answer. If its preview differs from the source, report the extraction problem.