Knowledge AI
⚠️ This feature is currently in public beta. We welcome feedback and may make changes based on usage and input. While it can be used in production, please do so with awareness of potential updates.
Last updated
Was this helpful?
⚠️ This feature is currently in public beta. We welcome feedback and may make changes based on usage and input. While it can be used in production, please do so with awareness of potential updates.
Last updated
Was this helpful?
Want to build an intelligent virtual assistant backed by a Knowledge Base you upload?
Our Knowledge AI feature along with the Q&A node uses RAG (Retrieval Augmented Generation) APIs to achieve just that!
AI Studio now allows you to upload or link Sources and use them as your Knowledge Base (KB) as an alternative to creating , or even using the node. Knowledge AI leverages to generate automated responses by using the uploaded sources.
The Knowledge AI functionality works through the Q&A node, however, there are some prerequisites to get the node to work, namely Sources and Indexes.
Let’s take a look at what these are and how to set them up within your account.
Step 1: Upload your Knowledge Base
Setting up this feature first requires you to visit the Knowledge AI tab after you log into AI Studio.
Sources are the first step in setting up your Q&A node's functionality. By uploading or linking Sources, you allow your VA to rely on tailored information suited to resolving your end users' queries.
What can you upload as a Source?
Local Files: In PDF, TXT and HTML formats.
All text from your uploaded sources, excluding table content or text embedded within images, can be used as a legitimate Knowledge Base.
Quick Note on using URLs
When URLs are uploaded, the content that is accessible on the direct site (seed URL, i.e main site that is visible upon clicking the link) is accessible by AI Studio.
This means that subpages are not automatically pulled and any updates made to the content are not automatically updated in the Source.
URL sets or Bulk URLs also need to be uploaded separately.
This means that domains and sitemaps will not work, each URL needs to be added individually.
Once successfully uploaded, the platform essentially takes a ‘screenshot’ of the URL in its current state and saves the text to use.
Please note that if your source is behind authentication walls or is inaccessible to the platform (eg: password-protected sites) you will run into errors in uploading. Uploading HTML files is a quick way to get around this.
File Size Restrictions
Your Sources are subject to the following restrictions:-
PDF file size limit: 10MB
HTML & TXT file size limit: 5MB
URL text size limit: 5MB
Need to update your Content?
It is imperative to remember that Sources are static, once uploaded/linked and processed, the content available within the Knowledge AI tab will remain unchanged. This applies to both local files and URLs.
It is important to re-upload/re-link sources when they go through content updates.
Ensure that the new Source is successfully uploaded before you delete the old Source and add the new Source to the concerned Index (Source group, more on this in the next step).
Additionally, you will have to name your Source. Please make sure to name your Sources in a way that makes the most logical sense for your VA so that you are able to accurately identify the right source at a later stage.
Where are your Sources accessible?
Sources are uploaded per API key and groups of sources (called Indexes, explained below) are accessible by VAs created within the relevant API key.
This means you can upload Sources to service different VAs under the same API key, however, any logical groupings of sources/Indexes created per individual VA need to be re-created manually if required for other VAs under the same API key.
For example, if we have:-
API key: D28326167
VAs under D28326167: VA 1, VA2
Sources uploaded under D28326167: Source1, Source2, Source3
Index 1: Source1, Source3
Index 2: Source2, Source 1
If VA1 uses Index 1 and VA2 also requires it, you will need to create a duplicate of Index 1 (and name it appropriately, eg. Index 1 Copy) in order to be able to use it in VA2, i.e each Index can be used multiple times but only within one VA.
You will know if your Sources are ready for use if the Status column displays a tick symbol against the Source.
The amount of time that your source may take to upload is dependent on the text size. This can range from about a minute to 30 minutes for extremely large sources.
Once successfully uploaded, you will also be able to preview any URLs that you have added, download your sources to your local files, and delete the uploaded sources on this page.
Step 2: Create categories for your sources
Once you upload your sources, you will need to create a group of Sources that your VA will use as a database within the course of the conversation, i.e. Indexes.
For example, if you have multiple sources related to troubleshooting a cooling mechanism for a specific model of CPU, you can group all of them under one Index so all questions related to that CPU can use these sources as a reference.
Similarly, if you only sell one type of CPU and have one main document to troubleshoot all issues, you will still have to put that single Source in an index in order for you to use it.
The same Source can be linked to multiple Indexes! This allows you a wider range of flexibility in terms of what your VA can answer in a single step.
Once created, an Index can be chosen in the Q&A node within the VA canvas to strategically provide the VA with the relevant information at the appropriate points in the flow.
Using Indexes in multiple VAs
All Indexes, like Sources are also stored at API key level, however each Index is limited to usage within one VA.
This means that if Index A is used in VA 1 it cannot be used in VA 2 even if VA 2 is under the same API key. However, Index A can be used multiple times within VA 1.
What do you do if you want to use the same Index in 2 different VAs?
You can create two Indexes with the same Sources so that you can use them across multiple agents. Be sure to name them appropriately so that you are able to identify them later!
This also means that deleting Sources and indexes will affect every VA they are being used in.
But before we get to actually using your Indexes and Sources within the Q&A node, how do you know if the index you created works for the right user queries?
Enter Index Testing!
Step 3: Test your VA for individual Index performance
Index testing allows you to see how each Index performs against anticipated user input.
In the event that Knowledge AI does not find relevant information, it will return “I don’t know” as a response. This is helpful in terms of preventing hallucinations, the generation of incorrect results and also to help you design for instances where the right response cannot be found.
For each response provided within the Index Testing mechanism, you can check out what Source within the index was used to provide the answer.
The results of the testing will also give you a quick view into what optimization and tweaks are required to allow for the right responses for each user query.
Optimization can range from editing the Sources within Indexes to editing the content within Sources to better match user queries.
Please note that any changes made to Index will immediately be reflected in the VAs live behaviour without the need to re-publish your agent!
Make sure that your agents are fully tested and ready for use before your agent is live to prevent any unexpected behaviour. In the case that you would like to test out changes, we recommend that you make a duplicate Index and test changes there before making final edits on your original Index.
Please note that each test/request made through the Index tester is charged separately. Please contact your account manager to learn more about pricing.
Now that you have set up the Knowledge AI tab, it’s time to set up your Virtual Assistant to take advantage of your uploaded and tested database. Click here for step 4!
URLs: - Websites for example:
Cloud storage links for Publicly Accessible Files that do not require access permissions. Eg,
Think of this feature as a dedicated for Indexes. For optimal live performance, we recommend testing various expected real-life scenarios to understand the behaviour and responses generated by the feature.
