Preparing to install IBM Knowledge Catalog

enableKnowledgeGraph: true

enableKnowledgeGraph: true

By default, the Neo4j graph database is installed during installation or upgrade. This graph database is required if you want to use IBM Manta Data Lineage as your lineage service.

useFDB: true is required only for new installations with MANTA Automated Data Lineage.

useFDB: true

Business-term relationship search is available with both types of graph database.

enableKnowledgeGraph: true

enableDataQuality: true

You must set both parameters.

If you enable this feature, you can work with the default enableModelsOn setting or change the setting. With enableModelsOn: gpu or enableModelsOn: remote, you can also work with a custom model instead of the default model:

customModelSemanticEnrichment

Specify a custom model to use with the gen AI based enrichment features. That model must be available in watsonx_ai_ifm.

If you use a custom model for gen AI based enrichment, make sure that you have the number of GPUs that is required for the selected model.

Important: For custom models, the accuracy of results might vary.

See Deployment mode.

enableSemanticAutomation: true
enableSemanticEnrichment: true

Deployment mode for the models that are used for the gen AI capabilities

This option determines where the foundation models run that are used with the gen AI capabilities in metadata enrichment.

cpu

This is the default setting.

A Granite model (granite3-moe:3b-instruct-q4_K_M) is started on the internal containers with CPUs. This model is only used for expanding metadata, and term generation and assignment in metadata enrichment.

Running models on CPU can result in significantly lower performance and reduced output quality, particularly for tasks such as name generation. CPU-based inference should be used with caution and is generally better suited for development, testing, or cost-constrained scenarios. For production use, GPU-based inference is strongly recommended.

If you use CPU models, results can be improved by making use of glossary and abbreviation support, and by combining LLM-based generation with approaches such as fuzzy matching.

• An instance administrator can upload abbreviation files to the semantic automation pods for global use in all projects. See Uploading custom abbreviation files for use in metadata enrichment.
• Project administrators can upload additional abbreviation files to projects and configure the metadata enrichment settings for metadata expansion. See Custom abbreviation files for name generation and Default enrichment settings.

To improve the performance of models on CPU, a cluster administrator has these options:

• Increase the resources by providing more CPUs and memory to the semantic-text-generation pods, for example:

  oc set resources deployment semantic-text-generation -n ${PROJECT_CPD_INST_OPERANDS} --limits=cpu=10
  oc set resources deployment semantic-text-generation -n ${PROJECT_CPD_INST_OPERANDS} --limits=memory=4Gi

• Assign a dedicated cluster node (worker) to the pods on which models with runtime are deployed. For more information, see Setting up a dedicated node for models running on CPU.

Important: If you are upgrading and want to continue to run the models on GPU, you must set enableModelsOn: 'gpu'. Otherwise, manually remove the previously used ibm/granite-8b-code-instruct foundation model and the ibm/granite-3-8b-instruct model that is deployed with the upgrade as described in Removing foundation models from IBM watsonx.ai™.

If you initially were working with models on GPU and switch to models on CPU later, you might also need to remove the ibm/granite-3-8b-instruct model. However, check that the model isn't used by any other service before you do so.

gpu

LLMs run in a local inference foundation models component (watsonx_ai_ifm). All generative AI capabilities are available in this mode.

Important: With this mode, the inference foundation models component (watsonx_ai_ifm) is automatically installed. This option requires at least one GPU to support LLM-based enrichment and LLM-based glossary generation. The default model is granite 3.1-8b-instruct.

For information about supported GPUs, see the Hardware requirements.

remote

LLMs run in a watsonx.ai instance on another on-premises instance of IBM Software Hub or on IBM watsonx™ as a Service. A connection to the remote watsonx.ai instance must be configured. See Connecting to a remote watsonx.ai instance.

Important: The required models must be available and running in the remote watsonx.ai instance. Setting up a connection to the remote instance does not start the models. If a required model is not available in the remote watsonx.ai instance, no connection can be established and any tasks that rely on those models will fail.

You must also set enableSemanticAutomation: true and enableSemanticEnrichment: true.

oc scale deployment semantic-text-generation --replicas=0

enableModelsOn: 'cpu'|'gpu'|'remote'

Enable computation of vectors for the metadata and the data during metadata enrichment. These vectors are used, for example, for generating SQL queries from natural-language text (Text to SQL feature).

This settings is also required if you want to enable the Text to SQL feature.

You must also set enableSemanticAutomation: true.

enableSemanticAutomation: true
enableSemanticEmbedding: true

Enable Text-to-SQL capabilities for generating SQL queries from natural language input. These capabilities can be used for creating query-based data assets, for example, for data products, SQL-based data quality rules, or in searches.

If you enable this feature, you must also set enableSemanticEmbeddings: true, and enableModelsOn: gpu or enableModelsOn: remote.

With enableModelsOn: gpu, the Text-to-SQL model is started in watsonx_ai_ifm. The default model is granite-3-8b-instruct.

customModelTextToSql

Specify the ID of a custom model to use with the Text-to-SQL capabilities. That model must be available in watsonx_ai_ifm.

If you work with a custom model for Text-to-SQL, make sure that you have the number of GPUs that is required for the selected model.

enableTextToSql: true

Enable LLM-based search for assets and artifacts across all workspaces.

enableAISearch: true