Capture inbox
What this page is — the front door for documents that exist outside Papers: PDFs, text files, images, voice memos and meeting recordings, uploaded by people or pushed in by automated sources, then classified and converted into real documents.
What it is for — so the agreements that arrived as attachments come inside the system, where the register, retention and search apply to them.
The problem it solves — agreements received as attachments never enter the register, so their dates, retention and search coverage are missing.
Route: /org/papers/capture · Permissions: Show the Capture Inbox sidebar entry. for the page;
Create documents from a type. to upload, mint sources and convert; Use AI drafting/extraction/Q&A for documents. to
classify.
1. What it is
Every item follows one path:
| Status | Means | Next |
|---|---|---|
| received (grey) | The file arrived; nothing has read it | Classify, Convert or Discard |
| classified (blue) | The AI read it and suggested a type and a summary | Convert |
| converted (green) | It became a Papers document, with the original file attached | Open document |
| discarded (red) | Rejected — the stored file is deleted; the row stays for audit | — |
How a file is read before classification:
| File | Read by | Works for |
|---|---|---|
| PDF, TXT | Its embedded text layer | Born-digital documents — invoices, contracts, letters |
| Scanned PDF, PNG, JPG, TIFF | Same — but a scan has no text layer | Refused |
| MP3, WAV, M4A, AAC, OGG, WEBM, MP4, MOV | Speech-to-text transcription | Recordings with discernible speech |
Images and scanned PDFs are accepted uploads, but there is no image OCR. Classify returns: no text layer was found in this document — it looks like a scan or photo. Image OCR is not available on this deployment: upload a text-based PDF, or paste the content as a .txt capture.
You can still Convert without classifying, picking the type yourself; the file is attached, but no text is extracted for AI to structure.
2. Why you would use it
- Nothing important stays outside the system. A counter-signed PDF emailed back becomes a document with an obligation register and a retention date.
- Automated feeds need no polling. A mail pipe or scanner script pushes files to a secret URL.
- Recordings become documents. A meeting recording is transcribed and classified like any other text.
- Rejections are auditable. A discarded row remains, showing what arrived and that it was rejected.
3. Step by step
Upload — drag up to 20 files onto the drop-zone, or click it.
Classify — press Classify (or Transcribe & Classify for recordings). The chip shows the suggested type code and a one-to-two-line summary.
Convert — press Convert. The suggested type is used if classified; you land in the new document's workspace with the file attached. Use Structure captured content there to fill fields.
Automate — Inbound sources → New source → name and kind → Create → copy the URL now — the token is shown once.
4. Field reference
Limits
| Rule | Value |
|---|---|
| Document formats | PDF, PNG, JPG, JPEG, TIFF, TXT — 20 MB each |
| Recording formats | MP3, WAV, M4A, AAC, OGG, WEBM, MP4, MOV — 100 MB each |
| Files per upload | 20 |
| Pages read | 100 (papers.ocr_max_pages) |
| Characters read | 400,000 (papers.ocr_max_chars) |
| Scan detection | Under 40 characters of text is treated as a scan (papers.ocr_min_chars) |
| Transcription timeout | 300 seconds (papers.stt_timeout_s) |
Classify and Convert
| Step | Uses | Refusals |
|---|---|---|
| Classify | papers_auto_classify — 450 credits est. · 1,800 max per file | capture item in status "converted" cannot be classified · no text could be extracted from this file · the scan message above |
| Convert | No AI; no credits | a document type is required — pass type_id or type_code, or classify the item first · suggested type code "x" does not resolve to an active document type in this scope… |
| Discard | — | converted capture items cannot be discarded |
Re-classifying replaces the previous suggestion.
Inbound sources
| Field / column | Holds |
|---|---|
| Name | Required — what a colleague will recognise |
| Kind | Email, Webhook / integration, Watch folder, Mobile capture app — a label only |
| Token | Shown once; only its hash is stored |
| Agent call | POST multipart form data, file field files, same formats and limits as upload |
| Ingested files, Last ingest | Health: a source that has not ingested for a week has probably broken at the sending end |
| Revoke | Kills the token at once; cannot be undone — mint a new source to restore |
5. Worked example
The accounts team receives supplier invoices by email and signed contracts from reception.
- They mint a source Accounts mailbox pipe, kind Email, and paste the URL into the mail-server script. Over a week it ingests 142 files.
- Most invoices are born-digital PDFs. Classify on one returns chip
invand "Tax invoice from Gatiro Retail for 45,900". Convert — the invoice is a document with the PDF attached. 450 credits. - Reception scans a counter-signed NDA as a JPG. Classify is refused: no text layer. They press Convert directly, choose Mutual NDA (India), and fill the fields by hand from the attached image.
- A sales manager drops
client-call.m4a. Transcribe & Classify takes two minutes and suggestsmeeting_minutes. Converted; the transcript feeds Structure captured content. - Three spam PDFs from the mail pipe are discarded; their rows remain, action-less.
6. The admin contract
| Must be true | Where | What breaks without it |
|---|---|---|
| The role holds Create documents from a type. | Role editor | No drop-zone, sources or Convert |
| The role holds Use AI drafting/extraction/Q&A for documents. | Role editor | No Classify |
| Active types exist for what arrives | Type Designer | Suggested codes do not resolve; Convert needs a type picked by hand |
| Senders send born-digital PDFs where possible | Your suppliers and scanners | Scans cannot be classified |
| The speech-to-text service is running | Operations | Recordings fail to transcribe |
| Source URLs are kept secret | Whoever configures agents | Anyone holding one can push files into your inbox |
7. Downstream
| When… | Then |
|---|---|
| An item is converted | A draft document exists with the file attached; the row points at it for good |
| A source is revoked | Its agent's next push is refused; captured items stay |
| An item is discarded | The stored file is deleted; the row remains |
8. Don't confuse this with…
| Attachments | Files on an existing document. Capture creates a new one |
| Briefcase import | A structured document from another Orbit environment |
| Intake forms | Documents from form submissions, no file involved |
| Zonal extraction | Reads fixed regions of a standard form on a type |
9. Troubleshooting
| Symptom | Cause |
|---|---|
| no text layer was found… | The file is a scan or photo — upload a text-based PDF or a .txt capture |
| Up to 20 files at a time | Split the batch |
| A recording failed to classify | No discernible speech, or it exceeded the transcription timeout |
| The suggested type does not resolve | No active type has that code in this project |
| An automated source stopped | It was revoked, or the agent's configuration broke — check Last ingest |
| A converted row cannot be discarded | By design |