Skip to main content
Nebula

Nebula Changelog

Release notes for Nebula — the document intelligence engine.

8 updates

Improvement

Encrypted Files, Orientation Correction & Template Formats

Password-protected uploads wait for a password instead of failing, and pages are turned the right way up before they are read.

  • A password-protected upload parks as “needs password” and waits at an unlock step, instead of failing as unreadable.
  • Pages are checked for orientation before OCR and corrected, with a backup model when the primary one fails.
  • Excel template formats (.xltx, .xltm) go through the synchronous fast lane.
  • Spreadsheets carrying images are routed to the full pipeline, so nothing embedded in them is lost.
  • The upload allow-list accepts every file type Specter handles.
Feature

Organizations, Retention Controls & Zero-Data-Retention Keys

Bill a whole organization centrally, and decide per API key how long — if at all — Nebula keeps what you send it.

  • Organizations: an owner, members, and billing held centrally for the organization rather than per user.
  • Zero data retention can be opted into when an API key is created — content is swept rather than kept.
  • A retention purge engine, plus a 30-day erasure sweep after an organization is closed.
  • Terms of service consent is captured at signup, with a re-consent gate when the terms change.
  • A budget alert at 80% of your monthly cap, and an expiry date shown per credit purchase.
  • Close your own organization from the portal, without going through support.
  • Custom prompts and crop policy can be set per account for organization tenants.
Feature

Public API Docs, Self-Service Keys & Unified History

Read the API reference, mint your own key, and search your whole conversion history rather than the page in front of you.

  • Public API documentation — a guide plus an interactive reference you can call directly from the page.
  • Mint, list and revoke your own API keys from the customer portal.
  • One History view across single files and batches, with search that covers your full history rather than the current page.
  • Search inside a batch, seeded from whatever you searched for in History.
  • merged.md follows the order you requested the documents in, so the output is reproducible.
  • Unsupported file types are refused at the front door, with a reason you can act on.
Feature

Reasoning Effort, Layout-Aware OCR & Faster Throughput

Trade speed against accuracy per job, and read each region of a page with an instruction chosen for that region.

  • Choose reasoning effort per job — maximum speed at one end, maximum accuracy at the other — and the choice persists for your session.
  • Layout-aware crop OCR: the page is segmented, each region cropped, and each crop read with an instruction resolved for that region.
  • Handwritten regions are routed to the model that reads them best.
  • Images take a fast path straight to the vision model, skipping the render stage they do not need.
  • A distributed rate governor keeps concurrency dense without tripping provider limits, with per-model rates.
  • Model coverage extended across current-generation GPT and Claude models, each with effort variants.
Announcement

Nebula is publicly available

Nebula opened to anyone with an account: turn Japanese and English business documents into layout-preserved Markdown and structured JSON, in the browser or over the API.

  • Charts, tables, forms, presentations and scanned pages convert with their structure intact.
  • Japan-hosted, with per-tenant isolation. Documents are not used to train models.
  • Available in the browser at nebula.ur-ai.net and over an authenticated API.
Feature

Accounts, API Keys and Prepaid Credits

A customer portal with real accounts, keys you control, and prepaid credits that stop rather than surprise you.

  • A customer portal with email accounts and multi-factor authentication.
  • Prepaid credit billing through Stripe: top up a fixed pack or a custom amount, with a hard stop when the balance runs out.
  • A pricing page with tiers and value-framed packs, plus referral codes redeemable after signup.
  • Billing is charged per stage, including JSON extraction, against a usage ledger you can export.
  • Read-only billing endpoints available with an API key.
Improvement

A Japanese Fine-Tune for Information-Dense Pages

Our own fine-tune for Japanese IR and financial material, served on dedicated GPU capacity and switchable against the base engine.

  • A Japanese fine-tune of an open-weight vision-language model, trained on Japanese investor-relations and financial material.
  • Served on dedicated GPU capacity, so throughput does not depend on a third party’s queue.
  • One command switches between the base engine and the fine-tune, so both can be measured on exactly the same inputs.
  • Per-page timeouts are configurable, so a single difficult page cannot hold up a batch.
Feature

Model Choice, Custom Prompts and Per-Job Feedback

Pick the model on the upload screen, and tell us what a conversion got wrong on the job that got it wrong.

  • Pick the model on the upload screen, from every model currently active.
  • A per-job feedback thread in a floating widget — the report stays attached to the job it is about.
  • Custom prompts managed from a central store rather than edited per deployment.
  • Presentations convert through PDF, so slide decks keep their layout.