Knowledge Base – Functions and resource management

Resources are the central sources for agents. They are part of the knowledge base and form the basis for the answers that the agents give in customer communication. This article describes how resource management takes place in the hub and which settings are possible.

1. Resource management
   1.1 Add and delete resources
   1.2 API – Connecting an external knowledge database
   1.3 Information retrieval interval
   1.4 Connect resources and agents
2. Description of functions – Retrieval Augmented Generation (RAG)
   2.1 Information retrieval
   2.2 Knowledge check
   2.3 Instructions
   2.4 Answer generation
   2.5 Data extraction
   2.6 Supported LLMs
3. Guardrails
   
4. Missing Knowledge
5. AI search in the knowledge base
6. Answer templates in the knowledge base
   
   
   

1. Resource management

The Knowledge Base menu item in the All Content section contains all the resources used by the specific agents and the standard AI agent.

1.1 Adding and deleting resources

Resources can be added individually for all file types, or via a bulk import for web pages.

There are two ways to add individual resources: either in the All Content section or via the RAG button. In both cases, clicking the Add Resource button opens an input form.

You then select the source type and add the resource by entering the URL or PDF/CSV file, or by creating the text document or question-answer pair. Instructions for uploading CSV files are described in this article. For all other source types, please refer to the instructions in this article.

The source is automatically enabled when added and can be disabled using the green toggle in the Used? column. This chapter explains how to assign specific sources to individual agents.

Bulk Import of web pages

The bulk import of URLs allows you to create all subdomains of a website at the same time. This speeds up the import process and ensures that all relevant subdomains are stored completely in the knowledge base.

The implementation is described step by step below.

1. In the input form, select Webpage as the source type and click the plus icon. The Expert Import URLs window will then open.
   
   
2. Select the Crawl Domain option in the input mask.
3. Enter the main domain in the Domain field.
4. Click Crawl to start scanning all subdomains of the specified website. This process may take some time.
   
5. In the results list, the desired subdomains can be selected by checking or unchecking the boxes. Once the selection is complete, click the Continue button to open a dialogue box.
   
6. Optional: Web scraping rules can be stored by activating the Include web scraping options toggle. These rules are applied to all URLs to be imported. Global settings can also be activated via incl. general options.

7. Confirm the import by clicking on OK. All selected subdomains are available as resources in the Knowledge Base.

 

Delete resources

If you wish to delete one or more resources completely, select them by ticking the box on the left. The Delete button will then appear, allowing you to permanently delete the resources. Alternatively, individual resources can also be deleted via the three-dot menu by selecting Delete resource.

1.2 API – Connecting an external knowledge database

In the moinAI Hub, it is not only possible to store resources such as websites, PDFs or CSV files in the knowledge base, but also to use articles from an existing knowledge database. How to connect an external knowledge database is described in this article.

1.3 Automatic update

Update options ensure that changes to the website are automatically transferred to the knowledge base, so that the AI chatbot always responds based on the latest version. There are two options available for managing the up-to-date status of stored resources: setting up regular, automated update cycles, and real-time polling for websites with frequent content changes.

The choice of the appropriate update method depends on the use cases in which the resource is required:

• Standard update interval: This method is suitable for static information (e.g. company profiles or general FAQs) that only changes every few weeks.

• Real-time update: This function is optimised for volatile content (e.g. news tickers, current prices, opening hours or availability). Here, the system checks with every request whether the stored content is still within the defined validity period. If this has expired, the website is immediately scanned again.

Both methods are configured in the Knowledge Base Management menu via the context menu of the relevant resource. Locate the desired web page resource in the list of all content. Clicking the three-dot menu at the end of the row opens the options, where you can select Freshness.

Set the default refresh interval

The regular refresh can be controlled in the upper section of the edit screen.

• Ticking the Refresh regularly box starts the automated process.

• The refresh cycle is set in the Refresh interval drop-down menu. You can choose from intervals of 7, 14 or 30 days.

Set real-time update

For time-critical information, the Real-time section within the same input screen is used. The number of websites for which real-time updating can be enabled is limited by the selected licence package.

The Enable real-time update toggle activates live website scraping. Under Validity Period, you can specify how long the cached content is considered up to date.

• Recommended validity periods

  • 30 seconds: Highly volatile content (e.g. live tickers)

  • 5 minutes: Standard; good compromise between timeliness and performance

  • 15 – 30 minutes: Content changes several times per hour

  • 1 hour: Content changes approximately once a day

• Additional option ‘Always include this source’: This toggle can only be used when real-time updating is active. If enabled, this specific website will be consulted for a response with every query from the RAG system, regardless of whether the system classifies it as thematically relevant. This setting is therefore only recommended for main pages whose information should appear in almost every response from the AI chatbot.

Detection and verification within the system

• Status indicator: In the Updated column of the resource list, the message ‘⚡ Real-time’ appears instead of a date.
  

• AI Playground: For test requests in the AI Playground, a real-time query is indicated by a yellow lightning bolt icon ⚡ next to the source URL. This icon only appears if the cache has expired and an actual live scrape has taken place.
  

1.4 Connect resources and agents

The agent and source are linked by selecting the desired agent via the drop-down menu in the top right-hand corner and activating the resource for the selected agent. Activation takes place via the toggle in the Used? column.

If a resource is no longer up-to-date, it must be replaced or deactivated in order to prevent outdated information from being displayed. The first step is to select the agent via the drop-down menu. It is then deactivated using the toggle in the Used? column.

If the resource is to be removed completely, it must be deleted as described above.

2. Description of functions – Retrieval Augmented Generation (RAG)

This section can be accessed via the RAG button under the Knowledge Base menu item.

The settings provided here can be used to control the responses of AI agents or generative AI. Options for controlling the content of generated responses are described in this article.

2.1 Information retrieval

Knowledge retrieval forms the basis for generating responses. Here, the AI agent determines which sources are relevant to a user’s query. New resources can be added using the Add Resource button. Further details are provided in this chapter.

The agent selects a resource in a two-step process.

1. Summary review: The AI agent analyses the summary of a source.

2. Detailed analysis: Only if the summary suggests relevance is the full content used to generate a response.

Clicking the eye icon next to a resource opens a preview of that resource. For websites and documents, an automated summary is also available there. By clicking Edit abstract, the text can be customised and, if required, the customised summary can be protected during automatic updates.

2.2 Knowledge check

As part of the knowledge check, the content of the resources is checked to generate answers. The AI agent only responds to the user's query if it contains sufficient facts to generate a meaningful answer. If the knowledge check fails, i.e. if there is not enough information, a “not understood” message is displayed in the chat. The knowledge check is a protection against hallucination and incompletely correct answers from the Generative AI. The setting is (de)activated by ticking the corresponding box in the moinAI Knowledge Base in the RAG → Knowledge check area. The setting applies immediately to the preview after saving. Publishing is necessary for the setting to also apply to the live environment.

2.3 Instructions

Instructions instruct the agent to include an informative addition such as “Information is provided without guarantee” or to follow other specific, content-related instructions. These are specific, content-related instructions and not instructions on the tone of voice/general communication guide lines. The instructions are inserted in the text field. After saving, the instructions are active in the preview, but not in the live environment. In order for the instructions to also apply to the live chatbot, it is necessary to publish the adjustments (Cloud-Button besides the RAG button).

2.4 Answer generation

There are various large language models (LLMs) to choose from. These models form the technical basis for text generation and each have different capabilities. Selecting a specific LLM therefore serves to ensure the best possible use case-related functionality of the AI agents. 

2.5 Data extraction

Data extraction makes it possible to save information. This stored information can then be used in subsequent actions. The data is, for example, e-mail addresses or customer numbers. Other data is conceivable depending on the use case. Extracting the data has a positive effect on the user experience. As soon as the setting option for data extraction is available in the moinAI Hub, we will provide information about the specific setting here and in the Help Center.

2.6 Supported LLMs

Various language models are available for response generation, including GPT-4o mini, GPT-4o, GPT-o1, Mistral Large, Gemini 2.0 Flash, Claude 3.5 Sonnet, and DeepSeek-R1.

3. Guardrails

Guardrails are guidelines and safety mechanisms that ensure AI agents only respond and act within defined limits. The following safety guidelines are always active:

• Content restrictions: Prevent the output of undesirable or harmful content, e.g. hate speech or confidential information.
• Topic restrictions: Limit the expert agent to specific topics to prevent misinformation or abuse.
• Reliability assurance: Detect and reduce hallucinations by using only verified information from the knowledge base.
• Illegible or nonsensical language: Detect and ignore requests that have no meaningful content, consist of random characters, or are intentionally incomprehensible. This prevents spam or malicious testing of the AI's limits.
• Request for system information: Blocks requests that aim to query internal system information such as the system prompt, internal instructions, or implementation details.
• Rule circumvention: Detects and prevents attempts to get the AI chatbot to ignore its security policies or remove previously set restrictions. This protects the integrity of the guardrails.
• Identity impersonation: Blocks requests that ask the bot to impersonate a real or fictional person or organisation. This prevents abuse such as phishing or identity theft.
• Inappropriate behaviour: Filters requests that aim to provoke the AI chatbot into inappropriate behaviour, e.g. insults, provocations or deliberate misconduct.
• Offensive or abusive language: Blocks requests that contain offensive, discriminatory, or aggressive language, even if it is only subtle or indirect.

Optional

• Competition protection: Blocks requests about competing products or translates the competing request into a request about your own product. This article describes how to activate the competition protection.
• Additional compliance check: Performs an explicit second check of the query. In this case, both the content guardrails and the competition protection are checked again. The function is activated by clicking on the compliance control activated toggle.
  

4. Missing Knowledge

If the AI cannot answer a question, it is either because the query was not understood or because there is no resource available in the knowledge base to answer it. In both cases, the element Not understood is executed in the conversation.

In the moinAI Hub, these dialogues can be viewed via the Knowledge Base menu item using the sidebar icon in the top right-hand corner. The Missing knowledge section opens.

This lists all queries where the chatbot thought there was information in the knowledge base.

Clicking on the speech bubbles opens the entire chat history. These indicate topics that are missing and should be added to the resources and included in the Knowledge Base (see this chapter).

5. AI search and filter

The Knowledge Base offers advanced search and filter options for efficient management of sources. The default view displays only actively used resources.

Filter

• The Used and Unused buttons control whether only active or deactivated resources are listed.

• The buttons below the search bar control which source types are displayed in the list. Multiple selections are possible.

Search: Name vs. Content

The search can be refined using a toggle switch.

• Name: Search specifically for the names, titles and URLs of the resources.

• Content: Searches the entire body text of the stored sources as well as the question-answer pairs. Search terms such as specific questions can be entered here. This is particularly helpful for terms and information not included in the title.

If a search is performed with the Used filter active (i.e. only resources that have been used), a yellow notification banner will appear if matching results are found in sources that are currently deactivated. Clicking Show will display these results.

6. Answer templates in the knowledge base

Answer templates can be used to integrate elements such as slides or URL buttons into the generated chatbot responses. This visual representation of content improves the customer experience and puts the focus on the products. This article describes how to create answer templates.