Frequently asked questions

How does the AI Konfuzio work?

Documents (image, PDF, email, web pages, RSS feeds, etc.) are sent to Konfuzio from input management services, scanning solutions or document management systems. As part of the capture process, OCR converts the written text contained in the images into machine-readable text data. This text is read, interpreted and translated into meaningful structured data by Konfuzios' AI. Then, this data can be customized and enriched with additional information in Konfuzio's smartview. Any additional information added manually is fed back to the Konfuzio AI to allow the AI model to continuously learn and improve. In order to detect missing / incorrect data or implement a VAT ID checker for example, customers can apply their own specific validation rules to the structured data in the validation step. Based on a comprehensive understanding of documents, Konfuzio can automate advanced tasks such as incoming mail. Our standard formats for data exchange are CSV, XML and JSON. These contain the extracted, validated and processed information that can be easily integrated into any system. The Konfuzio AI can constantly learn from new data from an ERP or CRM system and improve existing results
More about the procedure

Do I have to be a tech expert to use Konfuzio?

No, our AI can be trained conveniently via web interface. You don't need to be an IT expert to train your own model and use it on a daily basis.

What happens with my data?

Detailed information on the integrity, confidentiality, availability and purpose limitation of data processing can be found via the link below.
More about privacy & security

Is Konfuzio limited to the types of documents shown?

No, Konfuzio is a learning system and has already processed a wide variety of documents. On the homepage we present examples of exciting applications. Konfuzio offers our customers the possibility to train and improve daily the most different types of documents within the scope of their requirements.

How does Konfuzio differ from competitors?

Konfuzio belongs to the newest generation of solutions on the market which use the latest deep learning technologies and go beyond OCR extraction. Thus, various documents are read and further processed with customer-specific models.
However, Konfuzio goes a few steps further by making it possible for the software to understand and structure nested documents with different contents. Thus, Konfuzio automatically recognizes individual documents and the paragraphs they contain. Other solutions on the market use manual methods, e.g. Rossum AI uses a placeholder that can be inserted manually to separate individual documents in a file: https://rossum.ai/help/article/splitting-pdf-into-multiple-documents/

What if we already have existing automation processes?

Konfuzio is a component technology that we use to ensure easy integration of services into existing systems and software applications. Automating simple tasks that do not require individual knowledge or cognitive understanding is nothing new in the field of robotics (RPA). However, cognitive automation goes beyond configuring simple if-then rules. Like human reasoning, it requires intuition and judgment. To analyze and understand documents (like a human brain), Konfuzio combines several technologies such as word embedding, Recurrent Neural Networks (RNNs) for understanding text sequences and Convolutional Neural Networks (CNNs) for visual analysis of elements such as tables or paragraphs.

What are the technical requirements?

A quick and easy implementation is possible via REST API through the Konfuzio Cloud. Our cloud-based solution is hosted on Microsoft servers and can be accessed via HTTPS REST API. For Enterprise Edition, we also offer a multi-cloud solution on Kubernetes for maximum scalability. On the customer side, only a minimum of hardware is required when using the cloud.
For a productive on-premise deployment, we recommend a Kubernetes cluster or virtual machine with 16 CPU cores and AVX2 support enabled, 32 GB of RAM, and a Nividia GPU, with at least CUDA 10.1 and CUDNN 7.0. Smaller setups are possible, but can lead to performance limitations. Running without a GPU significantly increases runtime for extracting and training medium to large datasets. Supported operating systems are Ubuntu 16.04.3 - 16.04.5, 18.04; Red Hat Enterprise Linux 7.4 - 7.6; CentOS 7.4 - 7.6.
More about on-premise documentation

How long does it take to train the AI after new documents are labeled?

The training duration depends on the size of the dataset and the dataset itself (e.g. amount of text, number of pages), as well as the configuration of the labels and templates. For datasets with several hundred documents, we observe a typical training duration of 5-10 hours.

How can the file size of individual documents be prevented from being multiplied by the application of OCR, resulting in longer loading times?

The default configuration applies OCR to each document and creates a PDF file with text embeddings by rendering at 300 DPI for each document. The DPI can be set to a lower value. Alternatively, if you only have digital-native PDFs, you can turn off OCR and PDF rendering.

How often are there updates to the on-premise version of Konfuzio and how can learned AI model results be accessed after an update?

We are continuously improving Konfuzio. More about this in our Docs. The frequency of on-premise updates depends on the on-premise operator's preferences and organizational structures. AI models are stored in separate files within the configured storage. Konfuzio generally maintains compatibility with older models during updates. A separate backup of these files is possible.