24.6. Knowledge & Knowledge Bases

The Knowledge Assistant allows you to serve quick answers to your customers from a knowledge base of your choosing.

Configuring a Knowledge Assistant

Start by navigating to the Natterbox Settings page and clicking onto the Data Sources / Expertise tab under the AI Configuration section :

Now you can begin configuring your knowledge bases, to start just hit 'Create’.

How to define a knowledge base

We recommend that you create few knowledge bases and use the Parameter and Tag metadata to help segregate your data by type or area of specialisation - see more on meta data below.

1. Set a Name and Description

   1. i.e. Service Knowledge - All knowledge items to support our service queries

2. Click Save

Now you can begin populating your knowledge base, the next step is determining what knowledge material you wish to utilise.

Will you be using documents, websites or a mixture of both?

Document filetypes we support :

• PDF

• CSV

• JSON

• TXT

4. (Optional) Click Add Document to begin uploading documents

5. (Optional) Click Add Website to begin uploading websites

   1. You must supply the full website URL including the http:// or https://

By default we recommend to always supply the root domain for a website (i.e. https://natterbox.com not https://natterbox.com/products

If you encounter a data ingestion error because the site is to large, then you should look at using a smaller specific pathed part of your website (i.e. https://natterbox.com/products)

Note - if your website has localised paths (i.e. /us/en/products ) then we recommend that you add each one path as separate entries to ensure all pages are extracted (i.e. /products & /us/en/products )

6. Click Upload

Now you’re done, as soon as you click upload it will immediately save the material against the knowledge base for you and begin the process of extracting all the data from the chosen website or document.

You should see the status move into ‘Initializing extraction’

Note a standard 10-30 page website and a standard 5-10 MB PDF will only take a few minutes to successfully extract, a large website or document should not take any longer than 5-10 minutes.

If the extraction has been successful you should see a status of ‘Ready’ :

If for whatever reason the extraction has failed, you will need to try again, simply hit the refresh button or delete the object and retry the upload from scratch :

Meta Data and Knowledge base structuring

Purpose and Structure of Knowledge Bases

• Purpose: Organize each knowledge base around a clear purpose, such as data visibility (public vs. private) or organizational relevance (e.g., by department like Sales, Support, Engineering).

• Keep it Lean: Aim for fewer, larger knowledge bases rather than many small ones to ease management.

• Internal Structuring: Instead of creating separate knowledge bases for every data type, group related data within a single knowledge base and use parameters or tags to filter content during queries. For example, a "public content" knowledge base could contain both website and Q&A data, differentiated by a "ContentType" parameter.

Benefits of Structured Knowledge:

• More precise queries by assistants and agents, leading to faster response times.

• Minimized search data, leading to more relevant results for better responses.

• Reduced management overhead with fewer, more meaningful knowledge bases.

Best Practices:

• Separate major data categories (visibility, department) using different knowledge bases.

• Avoid over-segmentation; prioritize well-structured content within fewer knowledge bases.

• Use parameters and tags for internal classification and filtering.

• Configure assistants and agents to leverage these filters.

Parameters vs. Tags:

• Parameters:

  • Structured and Enforced: Provide a consistent way to classify content with predefined values (e.g., Privacy: Confidential or Audience: Internal).

  • Mandatory: Every document uploaded must have a value for each defined parameter, ensuring consistent classification.

  • Critical: Enable reliable filtering for search and AI queries, preventing unclassified content.

• Tags:

  • Flexible and User-Defined: Optional, non-enforced labels (e.g., pricing, onboarding).

  • Optional: Multiple tags can be applied, but their use is not mandatory, which can lead to inconsistencies or missed classifications.

  • Risks: Can be forgotten, inconsistent, and prone to user error, making content harder to find if solely relied upon.

Using Both: We recommend using both parameters and tags for the most effective approach. This combines the enforced structure and protection of parameters with the flexible, topic-based labelling of tags. For instance:

• Customer-Facing Pricing Sheet: Parameters (Privacy: Public, Audience: Customer) ensure external safety, while tags (pricing, sales, q3) allow for easy topic-based search.

• Internal Engineering Document: Parameters (Privacy: Confidential, Audience: Internal) protect access, while tags (architecture, deployment) help engineers quickly locate relevant information.

To define a Parameter do the the following from the Knowledge base settings :

1. Click ‘Add Meta Data‘ on the Knowledge base

2. Upload a Website or Document

3. Set the Parameter value against the Knowledge base before saving

To define a Tag you must first upload a document or website, then :

1. Define the name of Tag in the Tags field

   1. Alternatively if you have already defined your Tag you can select it from ‘Tags used in the knowledgestore’

2. Click ‘Add Tag’

3. Upload your Website or Document

Configuring the Knowledge Assistant App

Now your knowledge base is ready you can swap over the Routing Policies and begin configuring your Knowledge Assistant.

After creating a Natterbox AI component from the Routing Policies, simply select the Knowledge App.

1. (Optional) Define the goal of the Knowledge Assistant

   1. We recommend you use the default to ensure the Knowledge Assistant simply routes the call and doesn’t go beyond it’s scope - but this can be overridden to better fit your use case, see below for some examples

2. Set the initial message to the customer

   1. Treat the Knowledge Assistant like an agent

      1. Give them a name

      2. Provide the welcome

      3. State you’re serving the customers questions

         1. Example “Hi welcome to Natterbox, I'm Alice and I'm an AI, how can I help you today?”

3. Select the Knowledge base to use

4. Filter by Tag and/or Parameter

   1. This is optional but recommended as it will ensure your Assistants and Agents are optimised by ensuring they’re referring to only the specific data they need to

      

      

5. Set Similarity Threshold

   1. Filters results by accuracy to ensure relevant responses. Higher thresholds ensure better accuracy, but may reduce number of hits - we recommend to use the default of 30, but if you find the

6. Test the knowledge

   1. Click Test Query

   2. Provide a question which should be answered by the data with the knowledge base and click Test

   3. The AI will provide you with a response, the same response your customer’s would receive (without any behavioural modifications from a Persona)

7. (Optional) Domain specific knowledge*

   1. Use the Domain Context to give the assistant helpful background such as:

      1. A summary of what is in the knowledge base

      2. References to product names, etc

      3. Specific constraints around what the knowledge base cannot answer

   2. Otherwise the domain specific knowledge can be used alternative to a knowledge base

8. Set the message to play customer once the customer’s call has been serviced

   1. i.e. “Great, I hope I was able to help you today, goodbye for now”

   2. Depending on whether your business has a post call survey or not you may want to factor that into this message

9. Enable and configure Human Escalation - see here for more on this

Goal prompt examples

Conference Information

Your goal is to assist the Customer by answering their questions using a structured knowledge base. Do not begin answering until the Customer has provided their first question or input. To set a welcoming tone at the start of the conversation, open with a friendly, professional remark — such as a brief comment about the event, a helpful tip, or a polite offer of assistance — before inviting the Customer to share their question.

Product Support

Your goal is to efficiently assist Customers by providing accurate answers to their product-related questions, using information from the support knowledge base. Hold back from answering until the Customer has provided their first question or input. Start the conversation with a professional and solution-focused opening remark — for example, letting the Customer know you’re here to help resolve any product issues they may have. Wait for the Customer’s question before offering information or using any tools.

Internal HR

Your goal is to assist Employees by answering their questions using information from the internal HR knowledge base. Do not begin answering until the Employee has provided their first question or input. To create a welcoming start, begin with a professional and approachable opening remark — such as a courteous greeting, a reassurance that you are here to help with HR-related queries, or a brief mention of common HR topics you can assist with. Wait for the Employee to share their question before providing any information or using any tools.

Knowledge Assistant Routing Trigger

The Knowledge Assistant will trigger the exit node once it has determined that all customer questions have been answered. Typically, it will ask the customer for confirmation before exiting.

Fallback Behavior

The Knowledge Assistant will only fall through to the container’s exit node or the next application within the container if an internal failure occurs.

FAQ

How many knowledge bases can I create?

A maximum of five knowledge bases can be created.

What is the largest website I can upload?

The contents of the pages determine how large of a website, but we support storing up to 5 million characters which roughly translates to 900-1,800 pages if there’s an average of 500-1000 words per page.

My website status says ‘data too large’, what do I do?

This indicates the website chosen has reached the maximum allowed size.

You should look at using a smaller specific pathed part of your website (i.e. https://natterbox.com/products) to extract instead.

Note - if your website has localised paths (i.e. /us/en/products ) then we recommend that you add each one path as separate entries to ensure all pages are extracted (i.e. /products & /us/en/products ).

What is the largest document I can upload?

We support storing up to 5 million characters which roughly translates to 900,000 words.

If my website changes, will my knowledge base be automatically updated?

No, you will need to navigate into your knowledge and click on the refresh option again your website.

Will hosted files on my website also be ingested?

No, all files will need to be uploaded separately as individual documents uploads.

Will text from within images be ingested?

No, for both websites and documents, data ingestion is only valid for text.