Layout-aware parsing
Reading order, column breaks, merged cells, and footnotes are resolved before any model sees the text. A table spanning pages 12 to 15 comes out as one table.
Cortex 3 is reading 42 document formats
Documents in.
Evidence out.
Cortex reads contracts, filings, claim packets, and scanned forms. It returns the fields you asked for, in the shape you defined, with the page and region each value came from. When it is not sure, it says so and routes the field to a person.
- SOC 2 Type II
- Runs inside your VPC
- Your files never train our models
supply-agreement-q3.pdf
24 pages- Counterparty
- Alder & Voss Logistics
- Effective date
- 1 October 2026
- Liability cap
- 12 months of fees
- Auto-renewal
- Needs review
18 of 19 fields resolved
3.1 s for 24 pages
Illustrative interface — sample values for this template.
Reading day one of the backlog at
- Meridian Legal
- Kestrel Health
- Portside Freight
- Vantage Claims
- Union Grain Bank
- Halden Registry
01 — The gap
A model that reads well is not yet a system you can rely on
Most document pilots stall in the same place. The demo looks convincing on ten clean PDFs, then the real archive arrives: fax-quality scans, tables that break across four pages, amendments that contradict the original, a language switch halfway through an annex.
The output is usually plausible and occasionally wrong, and there is no way to tell which is which without opening the source. So a person opens every one anyway, and the time saved goes back where it came from.
Cortex is built around a single constraint: every value it returns must point back to the place it came from, and every value it is unsure about must be visible as unsure. Accuracy matters. Knowing where accuracy ran out matters more.
02 — Platform
Six things the platform handles end to end
Parsing, extraction, retrieval, and review live in one pipeline, so a correction made by a reviewer is the same record your systems read a second later.
Grounded extraction
Every field carries a page number, a bounding region, and a confidence score. Click a value in the API response and the reviewer lands on the exact paragraph.
Schemas you define
Describe the record you want in JSON Schema. Cortex validates types, enforces enums, and fails loudly rather than inventing a value that fits the shape.
Questions across a corpus
Ask across 200,000 documents at once. Answers arrive with the passages behind them, and an empty answer when the corpus genuinely does not say.
Review that closes the loop
Fields below your confidence threshold queue for a human. Corrections are logged, versioned, and fed straight into the next evaluation run.
Diffs between versions
Re-run on an amended document and get a field-level diff: what moved, what was added, what silently disappeared from the appendix.
03 — Model performance
Numbers, and the method behind them
Measured on a held-out set of 4,800 documents across six industries — half of them scans below 200 dpi, because that is what archives actually look like. The scoring script and the annotation guide ship with your account so you can reproduce or dispute any row.
| Task | Cortex 3 | Cortex 2.4 | OCR + general LLM |
|---|---|---|---|
| Field extraction F1 — clean PDFs | 96.4 | 93.8 | 88.1 |
| Field extraction F1 — low-dpi scans | 91.2 | 84.7 | 71.5 |
| Table structure accuracy | 94.0 | 89.3 | 76.9 |
| Citation precision (page level) | 98.1 | 95.6 | not supported |
| Abstention when answer is absent | 92.7 | 80.4 | 44.2 |
| p95 latency, 30-page document | 4.1 s | 7.8 s | 19.4 s |
+6.5 pts
F1 gained on degraded scans versus our previous release — the hardest slice, and the one that decides whether a backlog is workable.
1 in 14
Fields that reach the review queue at the default threshold. Raise the bar and more arrive; the choice stays yours, per schema.
0
Customer documents in any training run. Improvements come from licensed and synthetic corpora, stated in the model card.
Figures are sample data prepared for this template and do not describe a real product. A benchmark is a snapshot of one test set, not a promise about your archive — run the evaluation on your own documents before you commit.
04 — Workflow
From first upload to production in four moves
Most teams reach a reviewed, production-shaped pipeline in their second week. Nothing here requires a data science hire.
-
01
Point it at a source
Connect object storage, a mail drop, or the upload API. Cortex inventories the folder, classifies document types, and tells you what it cannot open before you spend anything.
-
02
Describe the record
Write the schema in the editor or paste JSON Schema you already maintain. Add a sentence of guidance per field — that plain-language hint does most of the tuning work.
-
03
Run, review, correct
Process a sample of 200 and open the review queue. Each correction is stored as a labelled example, and the accuracy panel updates as your reviewers work.
-
04
Ship it downstream
Push records to your warehouse, ERP, or case system over webhook or SDK. Every payload keeps its citations, so the audit trail travels with the data.
05 — Deployment
Run it where your documents already live
Cortex ships as a managed service and as a self-hosted image on the same release train. The self-hosted build is not a stripped-down variant — it is the identical pipeline, with the control plane pointed at your own cluster.
-
Single-tenant inference. Dedicated GPU pools per customer. No shared batching, no neighbour effects on latency.
-
Air-gapped mode. The self-hosted image runs with egress fully blocked. Licence checks are offline and telemetry is opt-in.
-
Pinned model versions. You choose when to move. Old versions stay available for eighteen months so your evaluations stay comparable.
-
Replayable runs. Every job records its model version, schema hash, and prompt bundle, so a result from March can be rebuilt in November.
Pipeline
Ingest & classify
42 formats
Layout & reading order
on-device
Cortex 3 extraction
live
Auto-pass
93% of fields
Review
7% of fields
06 — Security
Written for the people who have to sign off
Document platforms sit on top of the most sensitive material an organisation owns. These are the controls, stated plainly enough to forward to your security team.
Encryption
TLS 1.3 in transit, AES-256 at rest, and customer-managed keys on Enterprise. Keys rotate on your schedule, not ours.
Certifications
SOC 2 Type II and ISO 27001, with penetration test summaries published each half year. Full reports under NDA.
Access control
SAML and OIDC single sign-on, SCIM provisioning, and permissions down to the individual schema and folder.
Audit log
Every view, export, and correction is recorded with actor, timestamp, and document id. Streamed to your SIEM in near real time.
Retention
Set retention per project, from zero-persistence processing to seven years. Deletion is verified and certificated.
Residency
Pin a project to one region and nothing — including logs and embeddings — leaves it. Cross-region failover is off by default.
| Region | Managed service | Self-hosted image | Data leaves region |
|---|---|---|---|
| Seoul | Available | Available | Never |
| Frankfurt | Available | Available | Never |
| Virginia | Available | Available | Never |
| Sydney | Q4 2026 | Available | Never |
07 — In practice
Four backlogs we know well
The pattern repeats across industries: a queue of documents, a small set of fields that matter, and a team reading every page to find them.
Contract review
Renewal dates, caps, indemnities, and assignment clauses lifted from a master agreement and its six amendments — with the conflicts flagged rather than merged.
Claims intake
Photographed forms, hospital letters, and repair invoices arriving in one email thread, sorted into a structured claim file before an adjuster opens it.
Trade documents
Bills of lading, packing lists, and certificates of origin reconciled against the purchase order, with mismatched quantities raised as exceptions.
Regulatory filings
Disclosure documents read at publication, with the numbered paragraphs that changed since the last version surfaced to the analyst on duty.
08 — Research notes
We publish what did not work too
Short, dated notes from the applied research team. No press releases, no leaderboard victory laps — method, result, and the part we are still unhappy with.
- 12 Jul 2026 Confidence that means something Raw token probabilities made a poor review signal: the model was most confident exactly where the layout misled it. Calibrating against reviewer corrections cut wasted review time by roughly a third. Evaluation
- 28 May 2026 Tables that refuse to end Multi-page tables with repeating headers broke every heuristic we tried. Treating continuation as a learned decision rather than a rule lifted structure accuracy on financial appendices by 11 points. Parsing
- 9 Apr 2026 Teaching a model to say nothing Abstention is a capability, not a failure state. Training on questions whose answers were deliberately removed nearly doubled correct refusals, at a cost of 0.4 points of F1 we accepted. Grounding
- 17 Feb 2026 A negative result on synthetic scans Synthetic degradation looked convincing and helped almost nothing. Real fax artefacts have structure our noise model never captured, so we went back to licensing genuine archives. Data
09 — Field reports
What changed on the desks that use it
The citations are what sold our risk committee. Not the accuracy number — the fact that anyone can click a value and land on the clause behind it.
Our scans are twenty years of photocopies. Three vendors demoed beautifully and folded on the real folder. This one held, and told us honestly which pages it could not read.
We run it air-gapped, so evaluation was entirely on our side. The reproducible scoring script mattered more to us than any published benchmark.
10 — Pricing
Priced per page, because that is what you count
No seat tax on reviewers — the people doing quality control should not be the expensive part. Pages roll over for one billing period.
Pilot
For a single schema and a real sample of your archive.
$0.04 / page
First 2,000 pages free · no card required
- Managed service, shared region
- Three schemas, two reviewers
- Citations and confidence scores
- Community support
Scale
Most teamsFor production pipelines with review and downstream delivery.
$0.021 / page
From 250,000 pages a year · annual commitment
- Pinned region and model version
- Unlimited schemas and reviewers
- SSO, SCIM, audit log streaming
- Webhooks, SDKs, warehouse sync
- 99.9% uptime commitment
Sovereign
For self-hosted, air-gapped, or regulated deployments.
Custom
Annual licence · volume independent
- Self-hosted image in your cluster
- Customer-managed encryption keys
- Named solutions engineer
- Model card and evaluation review
Prices shown are sample figures for this template. A page is one rendered side; a double-sided scan counts as two.
11 — Questions
The ones that come up in every evaluation
Do you train on our documents?
No. Customer content is never used for training, fine-tuning, or evaluation, and that is written into the contract rather than a policy page. Reviewer corrections stay inside your tenant and improve only your own extraction rules.
What happens when the model is wrong?
It will be, sometimes. That is why every field carries a citation and a calibrated confidence score, and why anything under your threshold goes to a reviewer instead of straight into a system of record. Corrections are logged with the original value so you can audit drift.
Which languages and formats are supported?
Twenty-six languages at production quality, including Korean, Japanese, and Arabic, with mixed-language documents handled per page. Forty-two formats covering PDF, TIFF, common image types, office documents, and email archives with attachments.
How long does an evaluation take?
Plan for two weeks. Week one is schema design and a 200-document sample; week two is a scored run against your own annotations. We will tell you if your documents are a poor fit before you spend the second week.
Can we leave without losing anything?
Schemas, extracted records, citations, and correction history export as JSON and Parquet at any time, with no export fee. Nothing about the output format requires our runtime to read it later.
Do you replace our review team?
No, and we would be sceptical of anyone claiming otherwise. Cortex changes what reviewers look at: instead of reading every page to find four fields, they judge the handful of fields the system flagged. The expertise is still the point.
12 — Get started
Bring us a folder you have given up on
Tell us what you are trying to read and we will run a scored sample on your documents within a week. If Cortex is the wrong tool for your archive, we would rather say so early.
- Sales
- 1900-0000, weekdays 09:00–18:00 KST
- hello@webis.example
- Office
- 123 Webis-ro, Webis-gu, Seoul 12345, KR
What happens next
- 1 A solutions engineer replies with three questions about your documents — not a calendar link.
- 2 You send twenty representative files, including the ugly ones. NDA first if you prefer.
- 3 We return a scored extraction with citations, and an honest read on what it would take in production.
Request access
We reply within one business day, from a person.