numind_ai

What is NuMind?

NuMind is an AI company specialized in document information extraction; its core products are the NuExtract platform and the NuExtract3 multimodal model. These tools enable the conversion of PDFs, scanned documents, images, tables, and text into structured JSON format or Markdown format suitable for RAG.

In its early stages, NuMind offered primarily text classification, entity recognition, and low-code NLP model training. The current website focuses on document extraction and OCR; the offerings are listed under NuExtract3, NuMarkdown, APIs, and enterprise-specific deployments.

Core functions

• Structured extraction: Returns fields and hierarchical relationships based on a JSON template defined by the user.
• Multimodal understanding: Processing text, images, as well as combinations of text and images simultaneously.
• Convert documents to Markdown: Identify headings, paragraphs, lists, tables, and formulas.
• Template generation: Create an extraction structure based on natural language instructions or sample documents.
• Example learning: Adding a small amount of input along with the standard output improves performance on specific tasks.
• Multilingual processing: Performs extraction and OCR on files in various languages.
• Batch interface: Processes documents in the production environment via asynchronous tasks.
• Python SDK: Create projects, upload examples, submit files, and obtain results.
• Private deployment: Supports enterprise private clouds or on-premises environments.
• Custom models: Provide fine-tuning for scenarios requiring high usage or high accuracy.

NuExtract3 model

NuExtract3 is a unified visual language reasoning model with 4 billion parameters, capable of handling both structured information extraction and the conversion of images into Markdown format. The input can be text, single-page or multi-page images, and it is possible to choose between inference mode and non-inference mode.

• Parsed JSON can be generated based on the template and instructions.
• Read data from invoices, forms, contracts, receipts, and complex tables.
• Convert document images into hierarchical Markdown.
• Identify the table and use markers suitable for expressing its structure.
• Preserve special elements such as mathematical formulas and code.
• Handles long inputs and large structured outputs.
• Inference in one’s own environment can be carried out using public weights.

Structured data extraction

The user first describes the required fields using JSON Schema or platform templates, and then provides documentation along with optional examples. The model extracts information based on the target structure, making it suitable for storing unstructured files in databases, business systems, or within review processes.

• Invoice: supplier, date, tax amount, total amount, and detail lines.
• Contract: parties involved, deadline, amount, terms for renewal and termination.
• Resume: Name, experience, skills, education, and certificates.
• Insurance: policies, claims, losses, liabilities, and supplementary information.
• Healthcare: form fields, coding, medications, and medical records.
• Bank: customer verification, transaction, loan, and compliance fields.
• Technical data: components, specifications, parameters, and hierarchical relationships.

NuMarkdown and OCR

NuMarkdown is used to convert complex documents into clean Markdown format, with a focus on layout, tables, and RAG preprocessing. The current NuExtract3 version has integrated the Markdown conversion capability into a unified model; the old NuMarkdown repository can still be used to understand this technical approach.

• Restore the title and paragraphs to a clear hierarchical structure.
• Convert complex tables into Markdown or HTML format.
• Use LaTeX to express mathematical formulas.
• Handles scanned documents, image-based PDFs, and abnormal layouts.
• To preserve a readable structure for vector retrieval and knowledge bases.
• First, analyze the layout through reasoning, and then generate the final text.

Learning from templates and examples

NuExtract organizes extraction tasks into projects, which include templates, instructions, settings, and examples. Templates can be generated automatically from natural language descriptions, or they can be created by converting JSON Schema files and then further modified.

• Use an empty string to represent the text field that needs to be extracted.
• Use an array to describe the detailed items that may appear repeatedly.
• Nested objects are used to express hierarchical relationships.
• Provide examples of correct output to explain the meaning of the fields.
• Negative samples are used to test whether fields without values remain empty.
• Lock the projects that have passed verification to prevent accidental changes.

API and Python SDK

NuExtract offers asynchronous extraction interfaces based on API keys, as well as an official Python SDK. Developers can create projects, define templates, manage examples, submit text or files, check the status of tasks, and export results.

• Generate templates from natural language descriptions.
• Create, copy, import, and export extraction projects.
• Add or update examples for the project.
• Submit text or files for structured extraction.
• Extract the document content into Markdown.
• Check the status of asynchronous tasks via polling or streaming.
• Cancel tasks and manage project sharing status.

Which users are it suitable for

• Development team: Automatically writes PDFs and images into the business database.
• Data team: Prepares structured data for analysis, retrieval, and model training.
• Financial institutions: Handle loans, account openings, transactions, and document verification.
• Insurance companies: Review policy documents, claims, and loss-related materials.
• Medical institutions: Extract forms, medical records, and coding fields.
• Legal team: Batch identification of contract terms and key dates.
• RAG developer: Converts complex PDFs into well-structured Markdown.
• Researchers: Evaluating or fine-tuning visual language models for small documents.

Typical use cases

• Convert the invoices received via email into details in the financial system.
• Generate a comparison table of terms from hundreds of contracts.
• Read the identification document and fill in the customer verification fields.
• Convert scanned documents into a searchable Markdown knowledge base.
• Extract the claim amount and cause of loss from the insurance attachment.
• Parse resumes and enter them into the recruitment system using a uniform structure.
• Extract specifications, models, and compatibility details from the product manual.
• Process files that cannot be sent to public models in a private environment.

Product advantages

• The model is designed specifically for document extraction and OCR training, with objectives that are more focused than those of general chat models.
• The same model outputs both JSON and Markdown simultaneously, thereby reducing the need for multiple OCR processes.
• It supports text, images, and multilingual documents.
• A small number of examples can be used to adapt to corporate fields and formats.
• The API uses a pay-as-you-go pricing model, with no fixed monthly fee threshold disclosed.
• Multiple options are available, including open-source weights, SDKs, SaaS solutions, and private deployment.
• The public model has a smaller scale, which is suitable for companies looking to control inference costs.

Usage restrictions

• OCR and data extraction may still miss fields, misread lines, or fail to understand complex relationships.
• Some of the official performance figures are based on internal benchmarks, and retesting is required using the corresponding documentation.
• The fact that the output is in a parseable JSON format does not necessarily mean that the field contents are correct.
• Low resolution, rotation, obstruction, and handwritten content can affect the results.
• Very long documents and large tables increase the costs associated with input, output, and reasoning.
• There are no fixed public quotes available for enterprise private deployment and customization.
• Open-source models require GPU memory, an inference framework, and operational capabilities.
• The old NuMind functions are not fully compatible with the current NuExtract data.

API pricing

NuExtract3 currently uses a pricing model in which separate fees are charged for input and output tokens; users can start using it for free after registration. Lower prices are available for batch processing, large-scale calls, and fine-tuning, while customized pricing is offered for enterprise-level private deployments based on usage levels.

Text tokens are usually words or subwords; on average, one English word corresponds to about 1.3 tokens. Image tokens represent areas of approximately 32×32 pixels. Formatted documents such as PDFs are converted into images by default, so the actual cost should be determined based on the statistics of the task in question.

Comparison between SaaS and on-premises deployment

How to create an extraction project

1. Select a type of document whose fields are stable and whose manual results can be verified.
2. Prepare masked samples with normal, low-resolution, and abnormal layout.
3. Register on the platform and create a new structured extraction project.
4. Generate templates in natural language or import JSON Schema.
5. Clarify field types, repeated arrays, and nested relationships.
6. Upload a small number of samples and correct the standard output field by field.
7. Run the document verification with accuracy without involving any configuration.
8. Set business rules and manual reviews for key fields.
9. Generate an API Key before connecting to the testing environment.

How to integrate into the production process

1. Calculate the daily number of documents, the average number of pages, and the estimated Token cost.
2. Determine the file source, output system, and failure retry mechanism.
3. Use the official SDK to create asynchronous tasks and securely save the task IDs.
4. Poll the task status to distinguish between processing, successful, failed, and canceled.
5. Use Schema to validate the JSON structure and check for required fields.
6. Manual review is assigned based on confidence rules or business amount.
7. Record the model version, template version, and each modification.
8. Continuously sample and compare with manual results, while monitoring field-level errors.
9. Evaluate batch processing, fine-tuning, or private deployment before large-scale deployment.

Suggestions for accuracy verification

• The accuracy is calculated based on individual fields, rather than merely determining whether the entire document was processed successfully.
• Evaluate table rows, amounts, dates, and enum fields separately.
• Retain negative samples with missing fields to prevent the model from filling in content.
• Supports scan, photo, rotation, and multi-page attachments.
• Compare the model outputs with the manual gold standard item by item.
• Run regression tests after upgrading templates, examples, or models.
• Do not rely solely on the confidence levels reported by models for critical operations.

Privacy and corporate security

NuMind offers SaaS solutions as well as options for private enterprise deployment; industries such as banking, insurance, and healthcare can choose between a private cloud or an on-premises environment. Before processing real data, it is necessary to specify in the contract the retention period, the sub-processors involved, the uses for training, mechanisms for cross-border data transfer, and procedures for data deletion.

• During the testing phase, use masked or synthetic documents first.
• The API Key is stored in the server-side key system.
• Isolate development, testing, and production permissions by project and environment.
• Logs should not record complete identification documents, medical records, or financial files.
• Clarify the responsibilities for upgrading, monitoring, and ensuring security in private deployments.
• Establish processes for data export, deletion, and incident response.

GitHub and the open-source status

On its official GitHub repository, NuMind makes NuExtract, NuMarkdown, and the SDKs for its platform available to the public. The code for NuExtract, along with the models it uses, are licensed under the MIT license; however, the commercial SaaS versions, NuExtract PRO features, and the enterprise-hosted components are not part of this open-source offering.

Basic information

Recommendation score

4.7 / 5. NuMind is suitable for developers and enterprises that need to convert complex documents into JSON or RAG data in a reliable manner. It offers public models, low-barrier pay-as-you-go APIs, as well as options for private deployment; however, the accuracy in production still needs to be verified using actual business data samples.

Frequently Asked Questions

What is NuMind doing mainly these days?

Currently, structured JSON is extracted from documents using NuExtract3, and images or PDFs are converted to Markdown.

Does NuExtract support scanned documents?

Yes, it can handle image-based PDFs, scanned documents, and photos, and utilizes visual understanding to perform OCR and field extraction.

Can the output fields be customized?

Yes, fields, arrays, and hierarchies can be defined using JSON templates, natural-language descriptions, and a few proper examples.

How much does the NuExtract API cost?

The current rate is $1 per million tokens for input, and $5 per million tokens for output; discounts are available upon contacting the officials for large-scale usage.

Is the web interface free?

You can register and start creating and testing projects; production usage is billed according to the current token rules, with the specific amount of free credits depending on the account.

Is private deployment supported?

It supports private clouds and on-premises environments; custom adjustments are also available, and the price must be discussed with sales.

What is the relationship between NuMarkdown and NuExtract3?

NuMarkdown is dedicated to converting documents into Markdown format, while NuExtract3 has integrated structured data extraction and Markdown conversion into the same model.

Is a Python SDK available?

It provides functions for managing projects and examples, submitting text or files, as well as querying tasks to obtain results in JSON or Markdown format.

Is NuMind an open-source tool?

It is partially open source: the NuExtract model and the official SDK are made available publicly, but the hosting platform and enterprise services are not fully open source products.

Can it completely replace manual review?

It is not recommended; high-risk fields such as amounts, identities, medical information, and contracts still require rule-based verification as well as manual inspection.