IKS - Configuring Knowledge Collections

*See all Atlas AI & IKS articles here*

The Intelligent Knowledge Studio requires one or more Knowledge Collections to be configured in the tool to give users at least one option to chat with in the AI Assistant. In this article we explain how to create and configure these Knowledge Collections.

To perform these actions you will need to be part of the Atlas IKS Administrators Entra ID group.

In this article:

• Create and configure a Knowledge Collection
• Edit a Knowledge Collection
  
  

Create and configure a Knowledge Collection

Initial creation

1. To access the IKS go to the My Atlas menu under Settings:

   
   
   

2. From this panel you can add new collections, see the status for existing collections and edit them, and get reports on usage and user feedback:

   
   
   

3. To create a new Knowledge Collection click the New button at the top of the IKS panel. This will open a larger panel across the page where you configure the collection. First, set the title and security (permissions):
   
   You should make your title as descriptive as possible, especially if you expect to have a lot of collections. Next you need to choose the Owners, Auditors and Audience - you can choose individual users or groups - generally groups are recommended. You can click the (i) next to each security option to get a more detailed description.
   
   
4. Next, you need to set the Model Settings, including the type of model and the "system prompt" referred to here as Instructions. Detailed advice for this is not included in this article, but essentially you should tell the system in basic language what its role is and provide a brief on what users will ask and how to respond, including tone and the specific parts of the response required. The more specific your requirements for this collection the more specific your instructions should be.
   
5. The models shown in the configuration dropdown are subject to change over time as different models become available which may be more cost effective or more accurate. In general these will be labelled such as below where we have Advanced and Faster options. The models available as of Atlas 6.1 are as follows but please understand that this is a fast-moving area and models will likely change or be added to quite quickly:
   • "Faster" (GPT-4o mini);
     Optimised for speed and cost-efficiency. Ideal for scenarios requiring quick responses with moderate complexity. Suitable for real-time applications and tasks with budget constraints.
   • "Advanced" (GPT-4o);
     Enhanced understanding and generation capabilities. Best for complex tasks requiring nuanced comprehension and detailed responses. Suitable for knowledge-intensive applications and in-depth content generation.
   • GTP-4o mini for creative work;
     Used for creative work.
   
   The general advice is to start with the fastest and cheapest model, determine whether it is sufficient to serve your requirements, and if not to test an advanced model.
   
   
6. Finally we have the Grounding Settings. These values are where your AI expert can really drill into the results of testing and tweak the configuration of your collections to improve the quality of responses, and to maximise cost-effectiveness. This includes whether to use "strict grounding", maximum number of results to return for each request, and the required relevancy score: 
   

• The Strict grounding setting being turned on basically tells the model it can only use information from the Knowledge Collection. If you turn it off you essentially enable the assistant to act like a standard GPT with the ability to use it's own internal knowledge, but it will use the Knowledge Collection where appropriate.
• The Number of search results is defaulted to 50 - this basically limits how many results can be used to respond to the user. Increasing this would increase the potential cost of running a query, but would also potentially improve the scope of information available to respond. If you turn off strict grounding and reduce the number of results used to zero, you essentially now have a standard GPT which will simply act like any public GPT such as CoPilot, but with the model being held privately for your organisation.
• The Relevancy score setting basically determines how strictly the relevancy filtering is applied to the information coming from the Knowledge Collection. Essentially by lowering this score you increase the amount information the model will be able to confidently use to respond to users. By increasing the score, you may find the results are more accurate, but some queries may not be answered because the model cannot be confident enough about the information available.
  
  

Testing and modifying the configuration

The above options will need to be tested and tweaked for each Knowledge Collection to ensure your expected user scenarios are well supported and provide accurate responses - it is not something that can be determined in advance, and we recommend you test with the default settings first before tweaking.

Once you have completed the configuration on this screen, click Next at the bottom.

Set up the Knowledge Collection source

Once you set your model settings, we can now configure the source(s) for the model to use. You can select up to 1000 libraries from multiple workspaces. Use the interface to search for a workspace, then click the workspace to see the libraries, and select a library:

Already selected libraries are shown above the search, and you can click the X to remove them if you need to. Next, consider setting a KQL query filter. This allows you to use KQL just like in In Focus web parts, for example to filter results from the chosen libraries so that you only use information from the specified Entity:

You can click Test query to send the query to search. The system will simply report back the number of results:

Edit a Knowledge Collection

To edit a collection, from the IKS main panel, click the ... to the right of the collection and choose Edit. You can now make changes to all the configuration set in the previous section.