Collections and knowledge

A Collection is a set of documents an Agent can look up before answering. It is how the platform answers with your company's content, not just with what the model learned in training.

Collections belong to a workspace and are not shared between workspaces.

What a Collection is made of

A Collection brings together two kinds of sources, on the Files and Connectors tabs:

• Files: documents you upload, in PDF, DOCX, TXT and equivalent formats. They stay as they are until someone replaces them.
• Connectors: external sources the platform syncs, such as an indexed website or a document repository. They fetch content again from time to time.

A document's lifecycle

Every document goes through the same states before it becomes usable, and the Files tab shows the state of each one:

1. Queued: the file has arrived and is waiting to be processed.
2. Extracting: the platform is turning the file into text.
3. Indexing: the text is being split into passages and prepared for search.
4. Extracted: the document is available for answers.

Despite the name, Extracted is the final state: it comes after Indexing, not right after extraction.

The two error states are Extraction failed and Indexing failed. Neither recovers on its own, and a document in either state doesn't take part in answers.

Only Extracted documents are looked up. An Agent attached to a Collection whose files are still indexing answers as if the Collection were empty.

What happens on each question

When an Agent with Collections receives a message, the platform:

1. Converts the question into a numeric representation and searches for the closest passages in the Collections attached to that Agent.
2. Hands the passages it found to the model, along with the prompt and the history.
3. Asks for the answer, now with that material in view.

This design is what the industry calls RAG, retrieval-augmented generation. The practical consequence is that answer quality depends more on the indexed content than on the chosen model.

Limits worth knowing

• Search is by content, not by title. Asking for a file's name rarely finds the file. Use the terms that appear inside it.
• An empty Collection means "no knowledge", never "search everything". An Agent without Collections doesn't scan the workspace; it just doesn't search.
• Scanned PDFs depend on text extraction and take longer. When the original is an image without selectable text, extraction may fail.
• The scope is the workspace. Taking a Collection to another workspace means recreating it there.

Next steps

• See how the Collection connects to whoever answers in Agents and SuperAgents.
• Create your first Collection in Your first agent.