Classify
What this page is — the capability that reads a document and places it in your Papers taxonomy: its family, the best-matching type from your own catalog, filing tags, and the folder it belongs in.
What it is for — so an imported or badly-named file gets filed where a person would have filed it, and a backlog of untitled PDFs stops being a manual sorting job.
The problem it solves — imported files pile up untitled and unfiled, and are never found again by the people who need them.
Route: /org/papers/documents/<document>/ai → Understand → Classify ·
Permission: Use AI drafting/extraction/Q&A for documents. ·
Utility: papers_auto_classify · Cost: 450 credits estimated, 1,800 maximum.
1. What it is
Classify makes four judgements at once, and the important thing about all four is where the answers come from:
| Output | Chosen from |
|---|---|
| Family | A fixed list of seven — see §5 |
| Type | Your organisation's own document-type catalog |
| Folder | Your repository's existing folder paths |
| Tags | Freely generated from the content, lowercase, 3–8 of them |
The type code and folder path come only from types and folders that already exist in your organisation. When nothing fits, it returns an empty value rather than making one up.
So a document that keeps coming back with no type is not a failure of the classifier. It is the classifier telling you your catalog has a gap. Add the type or the folder, then re-run.
It judges by substance, not by the title. A file called scan_0043.pdf containing a tax invoice
classifies as a tax invoice. The current title is used only as a weak hint, and may be empty.
2. Why you would use it
Filing is the work everyone defers. The cost shows up months later, when the repository has four thousand documents and no reliable way to find the one you need.
- It makes an import tractable. A migration of legacy files is a sorting problem before it is anything else. Classification turns "3,000 PDFs in a folder" into "3,000 typed, tagged, filed documents to review".
- Tags are the part humans never do. Three to eight tags per document, drawn from subject, counterparty, domain and jurisdiction, is precisely the metadata that makes search work and precisely what nobody types by hand.
confidencelets you triage. Classify a batch, accept everything above your threshold, and spend human attention only on what came back uncertain.- It audits your catalog. A steady stream of empty
suggested_type_codeis evidence that a type is missing — a finding you would not otherwise get.
3. What you provide
Classify has two modes, and which one you are in is decided by where you run it.
| Mode | You provide | Use when |
|---|---|---|
| Document mode | Nothing | You have a Papers document open. It reads the body and title |
| Raw-text mode | The text, pasted | You are classifying an uploaded or legacy file whose content was extracted outside a Papers document |
Raw-text mode is the one that makes this capability useful for migration — it does not require the content to be a Papers document yet. Paste what the extractor gave you:
TAX INVOICE No. INV-2231 Bill To: Gatiro Retail... GST 18%... Total INR 45,900
The classifier does not need the whole document. A first page carrying the header, the parties and the totals is usually enough.
4. What it reads automatically
Uniquely among the AI capabilities, Classify reads your organisation's configuration as well as the document — it cannot suggest a type it has not been shown.
| Read | From | Used for |
|---|---|---|
| Document text | The open document | The classification itself, in document mode |
| Current title | The document | A weak hint only. May be empty |
| Available types | Your org catalog — code, name, family | The menu suggested_type_code is chosen from |
| Available folders | Your repository folder paths | The menu suggested_folder is chosen from |
5. What you get back
| Field | Type | Values | Means |
|---|---|---|---|
family | enum | transactional, agreement, correspondence, certificate, policy, record, custom | Exactly one. The broad class |
suggested_type_code | string | a code from your catalog, or empty | Best-matching type. Empty when none is a reasonable fit |
confidence | number | 0–1 | For family and type together, not separately |
suggested_tags[] | string[] | 3–8 entries, lowercase, no duplicates | Filing tags from subject, counterparty, domain, jurisdiction |
suggested_folder | string | an existing path, or empty | Best-fitting folder. Empty when none fits |
confidence covers the family and type jointly. A high score does not tell you the folder is right —
folder and tags carry no separate confidence, so review them on their own terms.
6. Worked example
A finance administrator is clearing an imported batch and runs Classify on a scanned file whose only name is a scanner serial number.
Input (raw-text mode):
TAX INVOICE No. INV-2231 Bill To: Gatiro Retail Pvt Ltd... GST 18%... Total INR 45,900
Result:
family:
transactionalsuggested_type_code:INVconfidence:0.94suggested_tags:["invoice", "gst", "gatiro", "fy2026"]suggested_folder:/Finance/Invoices/2026
What the administrator does with it. At 0.94 the family and type are accepted without review.
The folder is checked — /Finance/Invoices/2026 exists and is right — and the four tags are kept as
generated. Note fy2026: the classifier derived a fiscal-year tag from the invoice date, which is
the kind of tag that makes a later search work and which nobody types manually on document 300 of a
batch.
Had suggested_type_code come back empty, the finding would have been about the catalog rather than
the document: no invoice type exists to file it against.
7. Running it
- Open a document in Orbit Papers, or have the extracted text of a legacy file ready.
- Open the AI Assistant drawer, or go to the document's AI workspace.
- Choose Classify under Understand.
- In document mode there is nothing to type; in raw-text mode, paste the text.
- The family, type, tags and folder suggestion appear in the result tab.
The suggestions are not applied. Nothing about the document's type, folder or tags changes until a person accepts them. Classification is advice about filing, not filing.
Classify can be set to auto-run from the type's AI Config tab — though note the circularity: a document that already has a type does not usually need classifying. The useful case is the capture inbox, where documents arrive without one. See AI configuration.
8. The admin contract
| Must be true | Where | What happens if it is not |
|---|---|---|
| Your role holds Use AI drafting/extraction/Q&A for documents. | Role editor | The AI Assistant button does not appear |
The papers_auto_classify utility is active | Orbit AI Flow → utilities | "This utility is currently disabled" |
| The utility is enabled for your organisation | /org/ai-utilities | Absent from Understand, with no error |
papers.ai_monthly_credit_cap not yet reached | System Config | "monthly AI credit cap reached: n of n credits used this month" |
| Document types exist for the families you handle | Type designer | suggested_type_code comes back empty. Not an error — there is nothing to suggest |
| Repository folders exist | Repository | suggested_folder comes back empty for the same reason |
This is the capability whose quality depends most directly on configuration elsewhere. A rich catalog gets specific suggestions; an empty catalog gets empty fields and a family label.
For a confidential document, auto-run is default-denied — a platform administrator must set
papers.ai_confidential_cloud to the literal allow, and the block appears only in the server log.
A manual press still works, and where a local model key is configured the run is forced on-prem and
audited on the timeline as ai_local_routed.
9. Don't confuse this with…
| Type from document | Creates a new document type from a sample. Classify only picks from types that exist |
| Capture inbox | Where unfiled documents arrive. Classify is what tells you where they go |
| Smart fields | Extracts values from within a document. Classify decides what the document is |
| Zonal extraction | Reads fixed positions on a page. Classify reads meaning |
10. Troubleshooting
| Symptom | Cause |
|---|---|
suggested_type_code is empty | No type in your catalog is a reasonable fit. Add the type, then re-run (§1) |
suggested_folder is empty | No existing folder fits. It will not invent a path |
| The same wrong type keeps being suggested | Your catalog has no better option. The classifier picks the best available, not the right one |
| The document's type did not change | Suggestions are never applied automatically. A person accepts them (§7) |
confidence is high but the folder is wrong | confidence covers family and type only — folder and tags carry none (§5) |
| Fewer than three tags | The content is too thin to draw them from — common on a one-line letter |
| Raw-text mode is not offered | You are in document mode. Raw text is for content that is not yet a Papers document (§3) |
| The capability is missing from Understand | Role lacks Use AI drafting/extraction/Q&A for documents., or the utility is off for the organisation (§8). The document type does not gate it — AI Config controls auto-run only |