Blog

We have successfully integrated IBM watsonx Orchestrate with IBM Planning Analytics to allow seamless, real-time interaction with data and models using natural language for both end users and developers. This blog describes the initial proof of concept. We are developing functions to leverage this integration and deliver value-add functionality to end users as well as IBM PA model developers.

IBM watsonx Orchestrate is a central hub that helps coordinate workflows across various AI agents in an organization. In general, AI tools are designed to work with data found on the internet or stored in relational databases. Not many AI tools have the ability to connect to OLAP / multidimensional data sources. This integration establishes that connection, opening up possibilities for more effective uses in an analytical context.

The examples below focus on two use cases: i) technical application for IBM PA developers, and ii) business use case for application users.

IBM watsonx Orchestrate Integration for Developers

Our initial focus was to demonstrate the use of AI agent / LLM integration in support of IBM PA Development. The purpose of the integration is to execute various development and operational tasks using natural language, rather than writing code in IBM PA.  

Utilizing TM1’s REST API, we created a number of tool functions for the standard CRUD operations on the basic objects. We deployed these tools to our IBM watsonx Orchestrate (WXO) instance to give our agent access to our TM1 server. The initial list of tools is as follows: 

1. Ability to Create / Delete IBM PA Objects
   • Dimensions
   • Cubes
   • Processes
   • Rules
   • Chores
2. Read the Values and Content of IBM PA Objects
   • Dimensions
   • Cubes
   • Processes
   • Rules
   • Views
   • Chores
     
3. Write Content and Values into IBM Planning Analytics
   • TI Process Code
   • Rules
   • CellPut
   • Dimension Element Inserts
4. Miscellaneous Other Functions
   • Exists (All)
   • Run TI Process
   • Execute MDX

After spending some time debugging and testing to ensure the agent utilizes the functions correctly, we successfully enabled our WXO LLM to perform operations on the TM1 server based on queries in natural language. 

Here is an example query on WXO using the create dimension tool. The agent is using Claude Sonnet 4 from Anthropic as the base LLM.  

We can see that with simple prompting in plain English, an LLM can call certain tools, based on context, to accomplish defined tasks.  

Further, WXO allows the user to expand the logic to show reasoning steps for debugging and auditing purposes. An expanded view of the same query shows the tool logic and execution sequence. This is demonstrated below:

Building on this foundation, we continue to add tools to our WXO agent to expand its usefulness. For complex functions or functions with specific or difficult input formatting, it is crucial to provide adequate context for the model in the functions metadata. Additionally, functions should return a significant amount of context so that the model is prepared to give an adequate response, and to respond to follow-up questions. This design principle becomes particularly powerful when we consider how LLMs build upon previous chat interactions.  

Possible Applications: 

When starting up a TM1 instance on a new project, there is a plethora of prototype objects that are created manually (dimensions, cubes, views, and subsets) to confirm the overall design of the instance before starting the implementation. These are time-consuming to create and often require several revisions. The ability to quickly and accurately create TM1 objects based on descriptions in natural language would cut down on the initial configuration time. Instead of creating a custom TI process that will create dimensions and form a prototype cube from provided documentation, we should be able to feed any provided documentation to an LLM and describe the structure of the cube at a high level.  

In addition to prototyping, a TM1 developer would also be better equipped to handle errors and perform debugging operations if the LLM of choice has immediate access to TI code. Similarly to the popular offerings in major IDEs, granting an AI agent access to the code explicitly will give the user quick and easy access to AI powered insights on their code 

We will provide further updates as we develop more tools and functions. 

WXO Integration for Business Users 

The same integration can be leveraged to make information in IBM PA accessible to end users using the same AI agent and toolset. The complexity of the task is to recognize where the values in the IBM PA model reside and, where applicable, to differentiate rule-based outcomes and provide insight into the drivers. 

Here is an example of a simple query that pulls data from a specific intersection in a TM1 cube.

We see that the query not only returns the value of the cell in question, but also provides information about the formula and the value of its arguments. In this way, we are able to see the drivers behind the product sales, in addition to the product sales themselves.  

It is evident here that we did not explicitly query the model to provide any kind of trace for the value in question. However, as a result of our previous queries and contextual awareness, the model is capable of providing information about the location of the value (where the value came from), in addition to reading the requested cell from the cube. This demonstrates the ability of an LLM to provide context to the user by chaining together several tool calls. 

To illustrate this capability further, here is an open-ended prompt given to our agent: “Examine my cube [the same cube as above]”.  

This response called 10 different tools to gain an understanding of the structure of the cube. It also queried some data inside to develop a full picture. While the cube in question is a sample with limited data, this serves as a proof of concept that we are able to connect to the TM1 instance and effectively work with it leveraging natural language queries.  

Please reach out if you have questions or are interested in discussion. We will be posting more updates on this as we develop these capabilities further.