> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tryopencomputer.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Set up knowledge files

> Upload, process, attach, and verify reference material for an agent.

Use **Files** for reference documents the agent should consult, such as FAQs or policies. Use [Recordings](/guides/recordings.md) for reusable audio clips; those are a separate library.

## Upload and attach a document

1. Open **Files** and select or drop a document. The upload form accepts `.pdf`, `.docx`, `.doc`, `.txt`, `.json`, and `.md` files up to 5 MB.
2. Choose **Full Document** for small reference material that should be supplied in full on retrieval, or **Chunked Search** for larger material where relevant sections should be retrieved.
3. Select **Upload & Process**. Wait for processing to reach `completed`; `pending`, `processing`, or a stored filename alone does not mean the file is ready.
4. Open the target agent and edit the conversation nodes that need the material. Select the file under **Knowledge Base Documents**.
5. In the node prompt, say when to consult the document and how to respond when it does not contain the answer.
6. Save, test representative questions, and publish the workflow after checking the answers against the source document.

## Configure embeddings for search

Chunked/vector retrieval needs a compatible embeddings service in the selected model configuration. Configure its provider connection and embedding selection through [Models](/guides/models.md) before relying on search. Having a conversation model connected does not automatically supply embeddings.

If processing or search fails because embeddings are absent, configure the missing service or use an appropriate Full Document setup. The current deployment can return a generic server error for a missing embeddings selection; that error does not establish that the uploaded file is corrupt. An account without embedding credentials cannot complete an end-to-end vector-search check.

## Maintain the content

Open a document to inspect its processed content and the edit or processing actions available to your account. After updating content or retrieval mode, wait for any required processing to finish and repeat retrieval checks. Revisit agents that refer to a replaced or removed document.

Deactivating a document removes it from future use; it is not a receipt that the underlying stored object was physically erased. Keep a source copy outside the agent so your team can review changes.

## Verify retrieval quality

Use questions whose answers are definitely present, questions requiring multiple sections, and questions not answered by the document. Check that the agent uses the correct facts and declines to invent missing information. A completed index proves processing finished, not answer quality. Inspect the run for retrieval activity and check that the active node has the intended file attached.

For direct upload/process API sequencing, see [Tools and knowledge](/guides/tools.md#upload-and-process-a-knowledge-document).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.