Reference
Documents
Documents are your organisation's reference material: files you upload so agents can search and read them while they work. Once a document is processed it becomes part of the knowledge your agents can ground their answers in.
What it's for
The Documents area lets you give agents a shared, governed library of reference material: policies, product information, runbooks, contracts and anything else you want their work grounded in. Each uploaded file is processed for text extraction, then made available for agents to retrieve at run time.
Documents live under the Knowledge surface, alongside Memories. They are scoped to your tenant. Another organisation's documents are never visible, and a probe for one returns not found rather than disclosing it exists.
Uploading documents
Select Upload document, choose a file, and the upload begins immediately with a progress bar. There is no separate form to fill in — the file's name and size are read from the file itself.
-
Accepted formats: plain text
(
.txt), Markdown (.md,.markdown), HTML (.html,.htm) and PDF (.pdf). These are the formats the extraction pipeline can honestly read; other types are rejected at upload. - Size limit: up to the configured maximum (10 MB by default). The accepted formats and the current limit are shown beneath the upload button.
- One file at a time: each upload creates a single document; upload again to add more.
If a file is the wrong type or too large it is refused at the boundary and an error is shown — nothing is stored. On success you'll see a confirmation that the file has been uploaded and is being processed.
Processing and states
Every document moves through a fixed lifecycle, shown as a status badge in the list:
- Pending: uploaded and queued, waiting for processing to begin.
- Processing: text is being extracted from the file.
- Ready: extraction succeeded; the document is now searchable and readable by agents.
- Failed: extraction could not be completed. The reason is shown beneath the document name.
Ready and Failed are final states. Re-uploading the file is the way to recover from a failure. While any document is still pending or processing, the list refreshes itself automatically so badges settle without a manual reload.
You can narrow the list with the search box (by name) and the status filter, and page through larger libraries 25 at a time.
How agents retrieve them (RAG)
Agents never reach into this management surface. Instead, once a document is ready, its extracted text is split into chunks and embedded so it can be searched by meaning, not just keywords. During a run an agent uses read-only document tools to find the most relevant passages and read them (retrieval-augmented generation, or RAG) and weave them into its reasoning.
This means the practical way to make knowledge available to an agent is simply to upload it here and wait for it to reach ready. You don't wire documents to individual agents; the retrieval tools draw on your tenant's ready document library.
Deleting documents
Use the delete action on a row and confirm in the dialog. Deletion removes the stored file and the document's retrievable knowledge, so agents can no longer search or read it. This cannot be undone.
Governance and permissions
The entire Documents surface (listing, uploading and
deleting) is gated by a single permission,
documents.manage, held by the
owner, admin and
operator roles. Members without it cannot
see or change documents.
Uploads and deletions are recorded as domain events, so the audit log captures who added or removed each document and when. Because agents only ever read ready documents through dedicated tools, they never gain the ability to manage your library.