In the field of Intelligent Document Processing (IDP), Machine Learning (ML) models are increasingly being used to automate the process of extracting key-value pairs from different types of documents. These ML models are built by training on a set of labeled examples, where the goal is to learn to recognize and identify key-value pairs such as “document creation date” and “document number” within a document.

Once the model has been trained, it can be applied to new and unseen documents, allowing it to automatically extract relevant information such as dates and numbers. This extracted information can then be used to classify, organize, and manage documents effectively. The automation of these tasks, such as data entry and information retrieval, leads to significant time-saving and improved efficiency within the organization. Furthermore, the ML models in IDP are continuously updated and improved with new training data which allows them to adapt to changing document formats and increase accuracy.

But when it comes to extracting key-value pairs from documents, choosing the right model can make all the difference. While pre-trained models, such as those offered by large tech companies, are a popular choice for their ease of use, custom-tailored models may be a better option for achieving high accuracy and finding all key-value pairs.

Comparing Graip.AI’s ML Model with Microsoft’s Pre-Trained Model

In a recent test, we used a document in both English and Latvian to see how the two models handled the challenge. Both models performed well, but there were some noticeable differences in their approaches.

One of the main differences is that our custom model consistently adheres to the logic that all document text should be divided into key-value pairs. For example, the document title should be handled because it is often a document number with important explanatory information. However, the Microsoft model does not treat this information as important.

In general, the Microsoft model often ignores other important key-value pairs as well. This is not a good approach as splitting all text into keys and values is a better strategy. Even if our custom model makes an error, it can be corrected later on, but if the information is ignored, it is lost.

From the end-user’s perspective, using the Microsoft model can be less advantageous as important information may be lost. Additionally, we also found that the Microsoft model does not always correctly recognize key-value pairs.

While both models perform well, our custom ML model is better suited for extracting important information from documents. Its focus on dividing all text into keys and values and its ability to correct errors later on make it a more reliable choice for users looking to extract information from their documents. The Microsoft model on the other hand often ignores important information which can lead to losing important insights from the data.

Microsoft ML model

One of the main advantages of the Microsoft model is its ability to deliver good results out-of-the-box for documents in common languages, particularly English. The model is also well-supported by Microsoft, with updates being released every six months, which is quite good for such a large platform. Additionally, the model has a user-friendly interface and Microsoft invests heavily in AI research and development support. Microsoft also offers a limited ability to retrain its model, although it can be costly and difficult to do so.

On the other hand, one of the main drawbacks of the Microsoft model is that it can be considered outdated at any given time, which may not be convenient for users who have already integrated it into their system. Additionally, the model does not provide an interface for active learning and post-processing enhancements, and it can be difficult to predict the model’s results.

Graip.AI ML model

One of the main advantages of our model is its focus on classifying the entire document and searching for key-value pairs, ensuring that no information is lost. We have also demonstrated a very high quality of document recognition. Additionally, our ML model supports different languages and we can provide on-premise hosting, allowing users to place and store their data on their own servers for added security.

Another advantage of our model is its legal compliance, meeting all necessary security standards. With our model, users have full control over retraining and post-processing, allowing for active learning.

Our model also gives users full control over all versions of the model and allows for easy integration with third-party solutions.

In conclusion, our custom model offers a number of advantages over other models on the market, including its focus on classifying the entire document and searching for key-value pairs, its high quality of document recognition, its support for different languages, its on-premise hosting options, and its legal compliance. Additionally, users have full control over retraining and post-processing and can integrate it with third-party solutions.


What sets the Graip.AI model apart from others on the market is its focus on classifying all text in a document into keys and values and defining the relationships between them. This approach ensures that no information is lost and allows for a more comprehensive understanding of the data.

Another key advantage of the Graip.AI model is its active learning feature. This allows the client to mark up the data themselves and, after the model is trained, only make slight tweaks to the results if necessary. This results in a fully automated system where the customer manages their own cycles and templates.

In addition to its superior performance, the Graip.AI model is also more cost-effective than its counterpart from Microsoft.

Overall, the Graip.AI model represents a significant step forward in the field of Intelligent Data Processing (IDP). Its focus on classifying all text in a document, its active learning feature, and its cost-effectiveness make it a highly attractive option for businesses and organizations looking to extract valuable insights from their data.