The NeuroVision document recognition API takes a photo or a scan, detects the document type on its own and returns JSON: fields, a confidence score for each field and the results of anti-fraud checks. Connect the REST API to your backend or the Web SDK to your website — in the cloud in Russia or inside your own perimeter.
One request, one response. Your backend does not need to pick the document type, cut out the fields or calculate checksums: the service does it. Need to match a face to a document or check liveness? These are other steps of the same API — KYC and Liveness.
A permanent token is created in the personal account under “Access”; a temporary one — with a request to POST /user/login. Pass the token in the header or as the token field in the request body.
POST /kyc/process accepts one or more images inmultipart/form-data: for example, a passport spread and the registration page. The mode is synchronous or asynchronous.
The neural network finds the document in the photo, detects the country and type, straightens the image, splits it into fields and reads each one. Anti-fraud checks run in parallel: editing, a photo of a screen, a printout.
The response contains the document type, the country code, fields with confidence scores, check statuses and errors. In asynchronous mode the result arrives at your callback URL or on a status request.
Ordinary OCR returns a block of text, and your code has to parse it. The NeuroVision OCR API returns data that is already sorted out: each field with a name, a value and the model’s confidence score. Fields with high confidence can be written straight into a form, a CRM or a core banking system, and doubtful ones shown to the user or an operator. This is how the service works as an IDP system — read more on the page on document recognition.
The outcome of the check and the link between the result and your user or application.
status — idle, processing, success, failed, suspicious, exception, expiredsessionId — UUID v1 (timeuuid) of the sessionspent — total processing time in secondsisClientNew — the client’s first visit or a repeat oneclientKey — the client key, up to 36 characters; used to look up a sessionschemaId / scenarioId — the KYC schema (scenario) from the personal accountcallback URL — a webhook with the final session resultThe result for each image is a results[] element; only the set of document fields changes.
type — document, selfie, facesdocName — the document type, for example Passport (2006)countryCode — the country, ISO 3166-1 alpha-3 (RUS)isBackSide — reverse side or nottries — the number of recognition attemptsocr.fields[]title — the field namevalue — the valueconf — the recognition confidence scoreWhy a session may fail even when the fields have been read.
checks[] (fraud)isEdited — the document has not been editedisDisplay — not photographed from a screenisPrinted — the document is printed, not a photo of a screenisXerocopy — not a photocopyisMRZValid — the MRZ is validisFieldsModified, isStampsOk, isCropped and others — each check has status and captionerrors[] — for example [document] unknown document typeimageObjects[] — recognized objects: stamps, signaturesThe service detects the document type on its own. The integration needs just one thing — a legible image in the right format.
Passports, ID cards, driving licences and residence permits from 195+ countries and territories: Cyrillic, Latin and other scripts.
A passport spread and the registration page can be sent in one request or in separate scenario steps: the service merges the data in one session.
You do not need to specify the country or type: the model classifies the document from the image. For very similar documents the type can be set in the scenario.
The service accepts only JPEG and PNG. HEIC, WEBP and PDF are rejected with the error [general] bad image — convert them on your side before sending.
The face must be at least 30×30 pixels. No glare, heavy compression or cropped edges: the whole document must be in the frame.
By default the anti-fraud flags printouts, photocopies and photos of a screen. If your process needs scans, these checks are switched off in the scenario editor.
The Web SDK is a ready-made JavaScript widget. It opens over your page, asks for camera access, guides the user on how to photograph the document and sends the images to the API itself: instructions → capture in a frame → quality check with a hint → completion. The outcome arrives as a webhook to your backend; the successCb callback on the front end fires only on a successful outcome.
Mobile app: WebView instead of a native SDK. There are no separate libraries for iOS and Android at the moment. The widget is responsive and works with a smartphone camera, so in a mobile app it is opened in an embedded browser or a WebView. One limitation: the browser blocks the camera inside a cross-origin iframe, so the widget cannot be embedded in someone else’s iframe.
clientKey is encrypted on your server with the scenario key (the documentation example is AES-256-CBC on Node.js) and only then passed to the browser: working with clientKey, widget parameters. Appearance: light or dark theme (theme), a “frosted glass” effect over the page (blurHost), embedding inside a block (mountElementId), callbacks successCb, closeCb, finalizedCb, a “sticky” 15-minute session (stickySession), language selection. Without an integration — a direct link to the widget or one-time KYC links from the personal account, up to 1,000 at a time.
Cloud — images are processed on NeuroVision servers in Russia, launch comes after a token is issued and a scenario is configured, and scaling is handled by NeuroVision. Inside your perimeter (on‑premises) — on your servers, with no calls to external services; NeuroVision engineers install the Docker containers on your servers; scaling — your servers, with two servers and a load balancer for production.
Cloud — the internet is needed for API requests, and the models update automatically, without your involvement. Inside your perimeter — the internet is not needed to operate, only to activate the licence and for updates, which can be installed as offline packages; the models are updated by agreement: NeuroVision engineers connect over a secure channel and test after the update.
Cloud — payment for recognized documents or in packages; suits online services, fintech and marketplaces that need a quick launch. Inside your perimeter — a licence; suits banks, state-owned companies, healthcare and anyone with an isolated environment.
The core is the same — what changes is where the service runs and who maintains it. NeuroVision cloud OCR service suits a quick start; installation inside your perimeter — when documents must not leave the perimeter. The API functions are identical in both options: the same method, the same JSON.
CPU with 16+ threads and AVX2 (AMD Ryzen 9 3900X or equivalent), NVIDIA RTX 3090 24 GB or equivalent (A5000, RTX 4080), 64 GB RAM, SSD from 1 TB.
CPU with 24+ threads (Intel Core i9-12900KF or equivalent), 2× RTX 3090 (48 GB) or equivalent (2× A5000, A6000, RTX 4090), 128 GB RAM, SSD/NVMe from 2 TB, network from 1 Gbit/s.
You prepare the servers and access over SSH or a secure channel; NeuroVision engineers install the containers and check that everything works. Updates are by agreement: a new version is switched on only after testing.
Images and results do not leave the perimeter.
The internet is needed only to activate the licence and for updates.
Check them on your own documents and load. How to measure OCR quality — in an article on AI OCR accuracy
Processing time
On production servers a full document check — upload, recognition, field extraction and checks — takes 500–600 ms on average, on the demo stand — up to 5 seconds. The actual session time is in the field spent.
Synchronous and asynchronous mode
mode=sync — the result in the same HTTP response, convenient for a form. mode=async — returns sessionIdstraight away; the result comes as a webhook or on a request to POST /kyc/session/status. In the documentation examples the client timeout is 30 and 120 seconds.
Limits and packages
The plan sets no request rate limits: a package can be used up in a day or in a year. When a package runs out the service is not blocked — you simply buy the next one. We discuss the load at the start of the trial.
How long data is stored
Results are available by sessionId through the API and in the personal account; the retention period is set in the scenario. Original images are stored for 6 months and then deleted. Inside your perimeter you set the periods.
The NeuroVision cloud runs on servers in Russia
Under Federal Law No. 152‑FZ, recording, storing and retrieving the personal data of Russian citizens using databases outside Russia is not allowed (Art. 18, Part 5). The cloud OCR API does not breach this requirement, and with installation inside your perimeter the data stays with you.
Every request carries a token
A temporary JWT is issued with an expiry date; permanent tokens are created in the personal account with no expiry or with an end date, and permissions are defined by the user’s role. Tokens are stored as secrets and are never passed to the browser.
clientKey and webhooks are encrypted
clientKey for the browser is encrypted with the scenario key (AES-256-CBC in the documentation example). A webhook body can be encrypted with a secret key, and personal data need not be sent at all — only statuses.
You decide how long to keep it
The retention period for results is set in the scenario; original images are deleted after 6 months. Reports and exports are in the personal account.
No internet
With the on-premises option, images and results are processed and stored only on your servers.
152‑FZ
a Russian developer, software in the registry, No. 19972
[general] bad image — the file sent is not an image [general] too small image — the image is too small to recognize [general] no documents on image — no documents were found in the image [document] text fields not visible — the fields are not visible [fraud] mobile device — the photo was taken from a mobile device screen invalid schema (invalid parameters or no image), 401 access denied, 403 invalid user credentials. A recognition error arrives in an ordinary response: status: failed and the reason in results[].errors in the format [group] description. failed does not always mean “could not read”: the fields may have been recognized, but the session will not pass because of low confidence or an anti-fraud check. Look at errors and checks in each element of results. From request to your first JSON. A developer needs one evening; the rest are organizational steps on our side.
Leave a request — a manager will give you access to the personal account. Trial access lasts 1 month.
In the personal account open the “Access” section (Access Tokens) and create a token (set an end date if needed). Store it as a secret.
In the “KYC/AML” section create a scenario: which documents and countries to accept, which checks to enable, and the callback URL for webhooks. Copy its identifier.
Run the request POST /kyc/process following the method documentation with your token and scenario. One photo of a document is enough for the first time. You can first see how it works in the demo.
Write the fields to your system and handle failed and low confidence. For production, enable asynchronous mode and a webhook with encryption.
You pay only for recognized documents: per document or in packages, with no limits on the rate of use. The first month is a free trial.
Calculate the cost
The project «Development of a multimodal system using artificial intelligence "Neuromed" for managing the diagnostic processes of a clinic on the basis of the "Neurovision" computer vision platform» is supported by the Innovation Promotion Foundation within the federal project «Domestic Solutions»
The Neurovision computer vision platform is registered with Rospatent (certificate No. 2023661508) and included in the Russian Software Registry (entry No. 19972 of 13.11.2023)
The right to use the Platform is granted under a licence agreement concluded by Neurovision with its customers
Tell us which documents you need to recognize, in what volume and where the service should run — in the cloud or inside your perimeter. We will give you access to the personal account and help with your first request.
Can't put your request into words? Book a call and our team will contact you at a convenient time
Book a callThe company is included in the register of IT companies and the register of personal data operators of the Russian Ministry of Digital Development.
Address: Moscow, Presnenskaya Embankment 12 — Moscow City business centre, Federation Tower. Nearest metro stations: Delovoy Tsentr and Mezhdunarodnaya.
How to improve recognition quality and build a secure integration — on our blog.
Answers on methods, SDKs, Docker deployment, formats, speed, data and trial access
It is a REST API that takes an image of a document and returns the recognized fields in JSON. The service detects the country and document type on its own, reads the fields, checks the image for editing and for a photo of a screen, and gives a confidence score for each field.
POST /kyc/process at https://api.neuro-vision.ru/v1/. The request carries a token, one or more images in the images field and a scenario identifier. The status of an asynchronous session is requested with the POST /kyc/session/status method.
There are no separate server-side libraries: the API works over plain HTTP, so any language will do. The documentation has ready-made Python scripts.
There are no native SDKs for Android and iOS at the moment. The Web SDK works in mobile browsers and opens inside an app through a WebView, so document capture can be built into a mobile app too.
Yes. The service is installed inside your perimeter as a set of Docker containers, and NeuroVision engineers carry out installation and updates. After installation it works without the internet: access is needed only to activate the licence and for updates.
The minimum configuration from the documentation is a CPU with 16+ threads and AVX2, an NVIDIA RTX 3090 with 24 GB of video memory or equivalent, 64 GB RAM and an SSD from 1 TB. For production with a high SLA we recommend two servers with a load balancer.
Only JPEG and PNG. HEIC, WEBP and PDF files are rejected with the error [general] bad image, so they have to be converted before sending.
According to the documentation, on production servers a full document check takes 500–600 ms on average, on the demo stand — up to 5 seconds. The actual time of each session is returned in the spent field.
The plan sets no rate limits: a package can be used up at any pace, and when it runs out the service is not blocked. We discuss the expected load at the start of trial access.
In the cloud — on NeuroVision servers in Russia; with installation inside your perimeter — only on your servers. The retention period for results is set in the scenario, images are deleted after 6 months, and webhooks can be sent without personal data.
Yes, we give trial access for 1 month, with a personal account, a token and documentation. Leave a request at the bottom of the page.
Ordinary OCR returns a block of text, and you have to parse it. The NeuroVision OCR API returns fields with names and confidence scores, the document type and country, and the results of image anti-fraud checks.
