feat: add figure in markdown (#98)

* feat: add figures in markdown

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* update to new docling-core and update test results with figures

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* update with improved docling-core

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

---------

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

tests/data/redp5695.md

@@ -1,7 +1,16 @@
 Front cover
 
+
+<!-- image -->
+
 ## IBM Cloud Pak for Data on IBM Z
 
+
+<!-- image -->
+
+
+<!-- image -->
+
 ## Executive overview
 
 Most industries are susceptible to fraud, which poses a risk to both businesses and consumers. According to The National Health Care Anti-Fraud Association, health care fraud alone causes the nation around $68 billion annually.$^{1}$ This statistic does not include the numerous other industries where fraudulent activities occur daily. In addition, the growing amount of data that enterprises own makes it difficult for them to detect fraud. Businesses can benefit by using an analytical platform to fully integrate their data with artificial intelligence (AI) technology.
@@ -37,6 +46,7 @@ To learn more about these features, see the IBM z16 product page.
 Figure 1 on page 3 shows a picture of the IBM z16 mainframe.
 
 Figure 1 IBM z16
+<!-- image -->
 
 ## IBM z16 and IBM LinuxONE Emperor 4 features
 
@@ -45,12 +55,14 @@ IBM Z are based on enterprise mainframe technology. Starting with transaction-ba
 Figure 2 provides a snapshot of the IBM Z processor roadmap, which depicts the journey of transformation and improvement.
 
 Figure 2 IBM Z: Processor roadmap
+<!-- image -->
 
 The IBM z16 and IBM LinuxONE Emperor 4 are the latest of the IBM Z, and they are developed with a 'built to build' focus to provide a powerful, cyberresilient, open, and secure platform for business with an extra focus on sustainability to help build sustainable data centers. Although the z16 server can host both IBM z/OSfi and Linux workloads, LinuxONE Emperor 4 is built to host Linux only workloads with a focus on consolidation and resiliency. Depending on the workload, consolidation from numerous x86 servers into a LinuxONE Emperor 4 can help reduce energy consumption by 75% and data center floor space by 50%, which helps to achieve the sustainability goals of the organization.
 
 Figure 3 on page 5 shows a summary of the system design of IBM LinuxONE Emperor 4 with the IBM Telum™ processor. The IBM Telum processor chip is designed to run enterprise applications efficiently where their data resides to embed AI with super low latency. The support for higher bandwidth and I/O rates is supported through FCP Express cards with an endpoint security solution. The memory subsystem supports up to 40 TB of memory.
 
 Figure 3 System design of IBM z16 LinuxONE Emperor 4
+<!-- image -->
 
 The IBM z16 and IBM LinuxONE Emperor 4 servers are built with 7-nm technology at a 5.2 GHz speed. They consist of four dual-chip modules (DCMs) per central processor complex (CPC) drawer, each of which is built with two 8-core Telum processor chips that has "first in the industry" on-chip acceleration for mid-transaction, real-time AI inferencing, which supports many different use cases, including fraud detection.
 
@@ -59,12 +71,14 @@ Each core has access to a huge private 32 MB L2 cache where up to 16 MB of the L
 Figure 4 provides more information about the features of AI Accelerator integration with the IBM Z processor cores.
 
 Figure 4 IBM z16 on-chip AI Accelerator integration with IBM Z processor cores
+<!-- image -->
 
 The IBM z16 and IBM LinuxONE Emperor 4 server platforms are built with the hardware features that are shown in Figure 4 with addressing data and AI workloads in mind. Regardless of where the ML and deep learning (DL) frameworks are used to build and train data and AI models, the inferencing on existing enterprise application data can happen along currently running enterprise business applications. CP4D 4.6 supports Tensorflow and IBM Snap ML frameworks, which are optimized to use the on-chip AI Accelerator during inferencing. Support for various other frameworks is planned for future releases.
 
 Figure 5 on page 7 shows the seamless integration of AI into existing enterprises workloads on the IBM z16 while leveraging the underlying hardware capabilities.
 
 Figure 5 Seamless integration
+<!-- image -->
 
 ## What is Cloud Pak for Data on IBM Z
 
@@ -75,6 +89,7 @@ CP4D on IBM Z provides enterprises with a resilient and secure private cloud pla
 Figure 6 shows a solution overview of CP4D. The infrastructure alternatives are shown at the bottom, and they include IBM Z and LinuxONE. They all leverage Red Hat OpenShift. Common Foundational Services come next, which offer clarity throughout the data and AI lifecycle, that is, from user access management to monitoring and service provisioning. A high-level view of the services is shown in the middle section. The services have several different capabilities that span the AI hierarchy. The platform can be expanded, and it offers a seamless user experience for all distinct personas across the AI lifecycle, from data gathering through AI infusion.
 
 Figure 6 Solution overview of Cloud Pak for Data
+<!-- image -->
 
 We highlight the four main pillars that make IBM Z the correct infrastructure for CP4D:
 
@@ -135,6 +150,7 @@ Traditional ML models' power most of today's ML applications in business and amo
 Figure 7 on page 11 provides an overview of the components that are supported on CP4D on IBM Z. You can leverage Watson Studio for model building, training, and validation, and WML for deployment of the model. Eventually, applications can use the AI inference endpoint to score the model.
 
 Figure 7 Developing, training, and deploying an AI model on Cloud Pak for Data on IBM Z and IBM LinuxONE
+<!-- image -->
 
 In summary, here are some of the reasons why you should choose AI on IBM Z:
 
@@ -227,6 +243,7 @@ The key point here is that risk exists throughout the entire AI lifecycle starti
 For example, a business can start testing a model before production for fairness metrics. For this task, enterprises need an end-to-end workflow with approvals to mitigate these risks and increase the scale of AI investments, as shown in Figure 8, which presents a typical AI model lifecycle in an enterprise.
 
 Figure 8 Typical AI model lifecycle
+<!-- image -->
 
 Due to regulations, more stakeholders adopt the typical AI model lifecycle to protect their brand from new end-to-end risks. To ensure various aspects of both regulatory compliance and security, the personas that must be involved include the chief financial officer (CFO), chief marketing officer (CMO), chief data officer (CDO), HR, and chief regulatory officer (CRO), along with the data engineers, data scientists, and business analysts, who build AI workflows.
 
@@ -279,44 +296,54 @@ GLYPH<SM590000> Enterprises can develop AI models by creating and training model
 Figure 9 on page 16 shows the end-to-end flow for a remote AI governance solution.
 
 Figure 9 Remote AI governance solution end-to-end flow
+<!-- image -->
 
 To achieve end-to-end AI governance, complete the following steps:
 
 1. Create a model entry in IBM OpenPages by using CP4D on a x86 platform, as shown in Figure 10.
 
 Figure 10 Creating a model entry in IBM OpenPages
+<!-- image -->
 
 2. Train a model by using Watson Studio and by using development tools such as Jupyter Notebook or JupyterLab on CP4D on Red Hat OpenShift on a virtual machine on IBM Z, as shown in Figure 11.
 
 Figure 11 Training an AI model by using Watson Studio
+<!-- image -->
 
 3. Deploy the model by using WML on CP4D on Red Hat OpenShift on a virtual machine on IBM Z, as shown in Figure 12.
 
 Figure 12 Deploying an AI model by using WML on Cloud Pak for Data
+<!-- image -->
 
 4. Track the external model lifecycle by browsing through the Catalogs/Platform assets catalog by using AI Factsheets and OpenPages while using CP4D on an x86 platform, as shown in Figure 13. The external model (deployed on CP4D on Red Hat OpenShift on a virtual machine on IBM Z) is saved as a platform asset catalog on the x86 platform.
 
 Figure 13 External model
+<!-- image -->
 
 You can track the model through each stage of the model lifecycle, as shown in Figure 14, by using AI Factsheets and OpenPages.
 
 Figure 14 Tracking the model
+<!-- image -->
 
 You can see that the model facts are tracked and synchronized to IBM OpenPages for risk management, as shown in Figure 15.
 
 Figure 15 Model facts that are tracked and synchronized to IBM OpenPages on an x86 platform
+<!-- image -->
 
 5. Create an external model by using IBM OpenScale on the x86 platform, as shown in Figure 16.
 
 Figure 16 Creating an external model on an x86 platform
+<!-- image -->
 
 IBM OpenScale provides a comprehensive dashboard that tracks fairness, quality monitoring, drift, and explainability of a model. Fairness determines whether your model produces biased outcomes. Quality determines how well your model predicts outcomes. Drift is the degradation of predictive performance over time. A sample is shown in Figure 17 on page 21.
 
 Figure 17 IBM OpenScale dashboard that is used to monitor the external model
+<!-- image -->
 
 You developed and deployed the AI model by using Watson Studio, WML on CP4D on Red Hat OpenShift on a virtual machine on IBM Z, and end-to-end AI model governance by leveraging AI Factsheets, OpenScale, and OpenPages on CP4D on a x86 platform. Figure 18 shows end-to-end AI governance when using IBM OpenPages, AI Factsheets, and OpenScale.
 
 Figure 18 Final result: End-to-end AI governance when using IBM OpenPages, AI Factsheets, and OpenScale
+<!-- image -->
 
 ## Use case 2: Credit default risk assessment
 
@@ -335,6 +362,7 @@ Financial institutions can leverage AI solutions by using ML techniques to predi
 Figure 19 on page 23 shows a sample architecture about how to design and develop an AI model for credit risk assessment on IBM Z. An IBM WebSpherefi Application Server is used for handling in-bound transactions, and CP4D is used for AI model lifecycle management that includes building, training, and deploying the model.
 
 Figure 19 Architecture for credit risk prediction by using an ML AI model on IBM Z
+<!-- image -->
 
 A data scientist can leverage Watson Studio to develop and train an AI model and WML to deploy and score the model. In this sample architecture, the WML Python run time leverages the ML framework, IBM Snap Machine Learning (Snap ML), for scoring, can leverage an integrated AI accelerator at the time of model import.
 
@@ -349,6 +377,7 @@ We showed how IBM Z enable customers to use AI frameworks to detect credit risk.
 Figure 20 shows an architecture for predicting credit risk by using DL on IBM Z.
 
 Figure 20 Architecture for credit risk prediction by using DL on IBM Z
+<!-- image -->
 
 Data scientists can start creating and training a DL AI model by using a Jupyter Notebook instance and Watson Studio. Then, they can deploy the model by using WML on CP4D running on IBM Z, which provides an endpoint. Other applications, including the IBM WebSphere server, can produce credit risk results by using the model's endpoint.
 
@@ -385,6 +414,7 @@ One possible solution is to build and train a TensorFlow based DL model that lea
 Figure 21 provides a high-level diagram of a clearing and settlement use case for financial transactions that uses CP4D on IBM Z and IBM LinuxONE.
 
 Figure 21 Clearing and settlement use case for financial transactions by using Cloud Pak for Data
+<!-- image -->
 
 Here are the steps of the high-level process flow:
 
@@ -441,6 +471,7 @@ Remaining Useful Life (RUL) is the remaining time or cycles that an aircraft eng
 Figure 22 provides an overview of the inferencing architecture for the RUL of an aircraft engine when using IBM Z.
 
 Figure 22 Inferencing architecture on IBM Z
+<!-- image -->
 
 Because we are looking into data-driven model development, the data set of our target is the run-to-failure data of the engine. We are looking into a supervised learning problem, and we use regression techniques to learn from the data. DL techniques such as Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRU) are our choice because we are looking into a time series data set. TensorFlow or PyTorch frameworks are leveraged to create models. AI governance monitors the data and model drift to maintain the model quality throughout the model's life.
 
@@ -461,6 +492,7 @@ Client-side applications can invoke a REST apiserver that handles some preproces
 Figure 23 on page 29 provides a more in-depth view of the architecture of an AI-based predictive maintenance application.
 
 Figure 23 In-depth architectural view
+<!-- image -->
 
 In summary, consider the following points while developing an AI-based predictive maintenance application:
 
@@ -501,6 +533,7 @@ AI is the current "market trend evolution" in video analytics and advancing the
 S
 
 Figure 24 Architecture for AI-powered video analytics
+<!-- image -->
 
 Live camera feeds or recorded videos of an infant's movement are the inputs for a pose detection model. This video streaming data was stored in IBM Cloudfi Object Storage for image processing. Video data must be transformed into frames so that the infant's body poses can be detected. These post-estimation components of the pipeline predict the location of all 17-person key points with 3 degrees of freedom each (x, y location and visibility) plus two virtual alignment key points. This approach also embraces a compute-intensive heat map prediction of infant body posture.
 
@@ -620,6 +653,7 @@ IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of Intern
 
 The following terms are trademarks or registered trademarks of International Business Machines Corporation, and might also be trademarks or registered trademarks in other countries.
 
+
 | Db2fi IBMfi          | IBM Watsonfi   | Redbooks (log o) fi Turbon   |
 |----------------------|----------------|------------------------------|
 |                      | IBM z16™       | omicfi                       |
@@ -640,10 +674,16 @@ UNIX is a registered trademark of The Open Group in the United States and other
 
 Other company, product, or service names may be trademarks or service marks of others.
 
+
+<!-- image -->
+
 Back cover
 
 REDP-5695-00
 
 ISBN 0738461067
 
-Printed in U.S.A.
+Printed in U.S.A.
+
+
+<!-- image -->