Attachments and Knowledge

Knowledge and attachments both give the assistant material to base its answers on. But how that material is used depends on how you attach it. There are three ways to add files, and they behave differently.

Three types of attachments – different behavior

1. Files uploaded directly in a conversation (chat)

When you upload a file directly in the chat with an assistant:

• The file is parsed and its content is placed directly into the prompt.
• No chunking or embedding takes place.
• It is only available in that conversation — the assistant cannot search this file later.
• It works like copying all the text from the document straight into your message.

2. Attachments connected to the assistant

When you connect a file as an attachment directly to the assistant in the Space:

• The file is chunked and embedded (converted into vectors).
• RAG search is used — the assistant searches among the chunks when answering questions.

3. Documents in a Collection

When you create a Collection and upload files to it:

• The files are chunked and embedded (converted into vectors).
• RAG search is used — the assistant searches among the chunks when answering questions.

Which one to choose

• Use a file in the chat for a one-off document you want the assistant to read in full for the current conversation only.
• Use an attachment on the assistant or a Collection for material the assistant should be able to search across every conversation. Both use RAG, so the assistant retrieves the most relevant chunks instead of reading the whole document each time.
• To connect a Collection, the knowledge source must first be made available in the assistant’s Space.