Shelf Search Copilot Feature Overview





What is Shelf Search Copilot?

Shelf Search Copilot is an AI-based feature that leverages all the capabilities of Azure OpenAI and Large Language Models and customer’s content stored in Shelf Knowledge Management System (Shelf KMS) to augment or even replace standard search capabilities by automatically delivering accurate answers to agent and customer queries and questions. It provides ready-to-use answers, summarizes important details from the original content, creates refined questions and answers, and stimulates ease for agents, content managers, and users alike due to its functionalities. Currently, the Copilot works on content stored in Shelf KMS, but the future releases anticipate its efficient work with content synced from any external sources into Shelf KMS.



Preconditions

Before you can start enjoying all the benefits of Shelf Search Copilot, you need to enable it. For this purpose, when logged in to Shelf under your valid credentials, go to Admin Panel > User Groups > [USER_GROUP_NAME] > App Permissions and switch the Search Copilot toggle on.

Figure 1. Enabling Shelf Search Copilot in User Group settings

Once Copilot is enabled, you now should be able to see its icon on your Homepage in Shelf KMS. 

Figure 2. Verifying Copilot’s availability after enabling it

When the Search Copilot is enabled, make sure to provide access to the needed libraries in Shelf NexGen KMS.


Software requirements 

There are certain software requirements to be accounted for. Read below to learn them.
  1. Shelf KMS and associated web-based features require WebSocket for their proper operation. Make sure your organization's firewall does not block it as well as does not block Shelf's domains associated with the Shelf products. Contact your organization’s network administrators to handle this issue.
  2. Shelf website, Shelf KMS, and other products do not use or host any advertisement. With that being said, to get the best user experience and ensure a flawless operation of Shelf Search Copilot, disable any ad blocking tools or browser extensions you may have installed. If enabled, these tools/extensions may result in a decreased speed of operation or otherwise worsen your experience of Shelf products use.
Shelf’s domains may include the following ones:

api.shelf.io

static.shelf.io

api.ws.shelf.io

pdfpreview.content.shelf.io

pdf-viewer.shelf.io

thumbnails.shelf.io


Most frequent issues and their troubleshooting 

The most frequent issues you may encounter during the use of Shelf Search Copilot are no response from the Copilot and its very slow performance. 
These issues may occur when your organization’s network settings do not match the standard communication flow between the end-user device (your instance of Shelf KMS and Copilot) and Shelf’s servers. 
In this case you need to make sure the following settings are properly configured to fix the issues:
  • WebSocket protocol with wss:// is allowed
  • HTTP Upgrade requests to api.ws.shelf.io over Port 443 are permitted.

Once the above settings are adjusted, the communication between your Shelf KMS (and Copilot) and Shelf servers will occur as follows:
  1. The end-user device (your Shelf KMS instance) initiates the standard HTTP connection to api.ws.shelf.io
  2. Shelf’s server responds and, if the response is acceptable, the relevant Upgrade header is sent to switch the communication protocol from HTTP to WebSocket
  3. The WebSocket connection is established over the same route, that is using Port 443 for wss://.

If the communication flow is adjusted as detailed above, your Search Copilot is expected to work with a proper speed and efficiency.




What functionality does Shelf Search Copilot have?

In fact, Shelf Search Copilot has dual functionality, and you can read and find more details below.

Enhancement of Shelf Search capabilities

Shelf Search Copilot enriches classic Shelf search by automatically providing accurate answers to user queries based on content in Shelf’s KMS. It is triggered automatically when the user runs their search. However, if needed it can be triggered by users manually from Shelf Home Dashboard.

Figure 2. Viewing the search results and the operation of Shelf Search Copilot
Note that our Copilot processes not only content from local libraries but also from external sources synced through Shelf Content Integration Layer (CIL). However, it is only valid if such sources have been added and synced with Shelf. For receiving more information on Shelf CIL, its setup and specifics, contact your organization’s admin personnel or Shelf’s Customer Support team at support@shelf.io or from the in-app chat support within the Shelf platform.

Additionally to providing automatic answers during the search, Shelf Search Copilot provides references to an article (Gem) used as a source for the answer and where the user can find detailed information. Such references are implemented as links to the respective Gems.

Figure 3. Reference to the Gem used to generate the answer

When in the Copilot widget, the user can leave their feedback on the Copilot’s answer by selecting the appropriate icon—thumb-up or thumb-down. If the user leaves the positive feedback—clicks the thumb-up icon—they are further able to leave more descriptive feedback by entering their opinion in the feedback text field. If the user leaves the negative feedback—clicks the thumb-down icon—they can further not only leave their opinion in the respective field but also select an appropriate checkbox to specify what exactly they find negative in the Copilot’s answer.

Figure 4. Feedback functionality in Copilot

Moreover, the user can leave their feedback on the Copilot’s operation at any time. To do so, they need to click the More Actions menu button—(...)—in the upper right corner of the Copilot widget and select the Send Feedback option. Following that they can enter their feedback in the respective text entry field and click the Send Feedback button to send it to Shelf.

Figure 4a. Sending the generalized feedback on Copilot
All this feedback—specific and generalized—helps Shelf improve efficiency of our AI algorithms and quality of the content.
If the suggested answer is not sufficient or lacking some important information, the user can ask additional questions to refine their query and receive a more accurate response. It can be done in the Copilot widget through the Chat field.

Figure 5. Asking refining questions and receiving additional details
If the answer is suitable, the user can copy it for further use by clicking the COPY button under the answer.

In the case when the Copilot does not find what you search for, the respective message is displayed in the Copilot widget. If it occurs, you need to rerun the search using a different keyword or key phrase. Another option is to provide more context to the Copilot for a successful search.

Figure 6. Copilot response if no match has been found during the search

If, for some reason, the user no longer wants to wait for the answer or wants to ask another question, they can stop the answer generation process at any time by clicking the STOP GENERATING button.

Figure 7. Stopping the answer generation


         

Search Copilot and Shelf KMS filters 

The Shelf Search Copilot can work on the entire Shelf KMS database—search for the relevant answer using all the content in all the libraries and folders, content types, categories, etc. This is what occurs when the user runs a search in Shelf KMS: both Shelf KMS and the Search Copilot, which triggers automatically, search for an answer to the user’s question.

Figure 8. Viewing search results retrieved by Shelf KMS and Search Copilot

Once the Search Copilot offers an answer, the user can ask a follow-up question inside the original conversation with the Copilot. The Copilot, in its turn, understands that the question is within the original search context and offers an answer within that context.

Figure 9. Copilot understands original context and provides accurate answer 

The Shelf Search Copilot is also capable of searching for answers within the content filtered by the user using Shelf KMS filters (library, folder, etc.). 

Figure 10. Viewing Search Copilot’s answer found in the content after Shelf KMS filters
The user then can ask follow-up questions within the filtered content and receive refined answers from the Copilot.

Figure 11. Viewing Copilot’s refined answer found in the filtered content


            

Running new search from ongoing Search Copilot conversation 

The Shelf Search Copilot is capable of understanding that a user’s question is the new query and run the new search and provide the new answer.


Figure 12. Running a new search from the ongoing conversation with Search Copilot




Enhancement of Shelf Gem capabilities

This part of the Shelf Search Copilot functionality improves the way users interact with content (Gems) stored in Shelf. The only thing the user needs to do is to launch Copilot from the Gem page. The supported Gem types are Article, Wiki, Post, FAQ, Document, and Image. In the case of the Image type of Gems, we mean text parts - title, description, associated wordings - of such Gems.

Figure 13. Viewing and working with Copilot on the Gem page

1) Summarization: A significant reduction in the time needed for the user to familiarize themselves with the Gem’s content, by showing a brief Summary of the Gem in the Copilot widget. This summary contains a compressed overview of the Gem’s content, making it possible for the user to learn about quickly, without having to read the entire Gem. Naturally, the user can copy this summary for further use by the simple button click. Whether the user finds the summary good or bad, they can leave their feedback by clicking the respective feedback icon.
2) Questions and Answers: The Copilot analyzes the Gem and creates a set of questions and answers that cover all the important aspects of the Gem’s content. Each answer can be copied from the Copilot widget for further use.
3) Refine: A user can ask additional questions to refine information about the Gem briefly presented in Summary and Q&A simply by entering their question in the Chat field and waiting for the answer.

Figure 14. Asking additional questions via Copilot Chat

While the answer is being generated, the user can stop the generation process by clicking the respective button (See Figure 7 above). 

The answer they get can be copied for further use. The user can also leave their feedback on this new answer as well, by clicking the thumb-up (if the answer is useful)  or thumb-down (if the answer isn’t useful or helpful) icon. It allows us to improve our AI algorithms and the content quality.

If no answer is found by Copilot in the subject Gem, the respective message is displayed as shown in the figure below.

Figure 15. Copilot response if no information has been found in the Gem


What are known limitations of Shelf Search Copilot? 

Along with the vast functionality, the search-enhancing component of the Shelf Search Copilot has the following limitations:
  • English only: no other languages are currently supported
  • No images in the output: only text is currently supported in the Copilot output
  • No direct reference to a specific Decision Tree step: though the Decision Tree Gem type is supported, Shelf Search Copilot currently does not provide pin-pointed references to specific steps of such Gems
  • No searches/answers are saved/stored: after you run a search with Copilot and get search results, you need to copy them to save for further use; the same applies to conversations with Copilot (questions and answers) - they are not saved and disappear when you click other links or open other documents. This limitation may be withdrawn in future releases of the Search Copilot but at present, if you want to save this data, you need to copy it or take screenshots of it.