Collections
A named pile of documents Asteria can search, and the single most useful thing you can set up in your first week.
A collection is a named set of files that Asteria can work with in any conversation, for as long as you keep it.
Why bother
Attaching files to a chat works, and it stops working the moment you have more than a handful or you want them again tomorrow. Attachments belong to one conversation. A collection belongs to you.
The difference in practice:
| Attach to a chat | Put in a collection | |
|---|---|---|
| Available in other conversations | No | Yes |
| Sensible number of files | A handful | Hundreds |
| Largest file | 50 MB | 250 MB |
| How a big document is handled | Read end to end, and it may not fit | Split up, searched, and the relevant parts cited |
| Shareable with colleagues | No | Yes |
The fourth row is the one people underestimate. A 300 page PDF in a collection gives better answers than the same PDF attached to a chat, because Asteria retrieves the passages that bear on your question and tells you where they came from, instead of trying to hold the whole thing in mind at once.
What goes in one
Anything you or your team refer back to. The pattern that works is one collection per body of knowledge, not one per project:
- Supplier contracts
- Everything the company has published about the product
- Board packs, quarter by quarter
- Your brand guidelines, tone of voice and past campaigns
- The research you have accumulated on a market
Two kinds of file, and they behave differently
This surprises people, so it is worth knowing on day one.
Documents (PDF, Word, PowerPoint, text, images) are read, split into passages and indexed. Ask a question and Asteria finds the relevant passages and cites them.
Data files (CSV, Excel, Parquet) are not indexed and are not searchable. They are stored, and when you ask a question about one, Asteria loads it and computes the answer.
That is deliberate, not a gap. The answer to a spreadsheet question is almost always an aggregate ("what did we spend in total", "which region grew fastest") that no single passage of text contains. Searching for fragments of a spreadsheet would find you a row; computing over it gets you the answer, and shows the working.
The practical consequence: do not expect a keyword in a spreadsheet to turn up in a search. Ask the question you actually want answered instead.
Where to go next
Creating a collection
Two fields and one decision that is worth thirty seconds of thought.
Adding documents
Formats, limits, what "Ready" means, and how to tell when something failed.
Asking a collection questions
The one rule that decides whether this works at all: say the name.
Letting Asteria file its own work
Deliverables saved back into a collection, revised in place, kept in one copy.
Connected folders
Point a collection at a Google Drive or SharePoint folder and let it keep itself up to date.
Sharing your work
Who can see a collection, and the one sharing decision you cannot undo.