Check and correct queries and analyze generated responses
If the chatbot's AI assigns a user enquiry to the wrong intent or no intent at all, or generates an inappropriate answer or no answer at all, it is helpful for the AI to receive feedback from an expert. At this point, it is important to show the AI what the correct intent would have been or what the appropriate answer would have been.
In the AI Playground menu item, the AI can be tested with queries and then optimized.
The AI Playground is an extended version of the former “AI Feedback” tool and offers a compact user interface with multiple functions & transparency to optimize the artificial intelligence of the chatbot.
How the AI Playground works
The first step is to enter a sample user request.
The AI checks whether a suitable editorial intent or agent exists and suggests it. The title, the scenario and the current status of the detected intent or agent are displayed. If neither a suitable agent nor a suitable editorial intent is recognized, a corresponding message is displayed.
The Edit button can be used to switch directly to the editor of the corresponding intent or agent in order to make adjustments.
The duration of recognition is displayed in seconds at the bottom right. You can reset all entries in the Playground by clicking the Reset button at the top right or by reloading the page.
Generate responses (AI agents)
With the Generate response button, the AI creates a response as it is also output in the live environment. The time taken by the AI to generate the answer is displayed on the left below the answer.
The display of the relevant documents can be expanded on the right below the answer. This list helps to understand which resources were considered by the AI for this answer and which specific partial information was used (button see content behind each resource).
✔️ Resource was used for answer generation.
X resource was considered for answer generation but not used.
The type (website, PDF, list, etc.) is also displayed for each resource and whether it can be accessed by users or is not publicly available, depending on the resource type.
⚠️ No answer could be generated. Not enough knowledge in the knowledge base.
If this message appears, no answer is generated in the playground or in the live environment. Once sufficient relevant information has been assigned to the agent via the knowledge base, an answer is generated. If the agent in question uses the knowledge check in the knowledge base settings, it can help to deactivate it so that a response is generated.
To ensure the best possible recognition by the AI chatbot, you should always check whether the request variant matches the intention of the stored intent.
Giving feedback
Feedback can be used to continuously improve the AI, leading to more automation, more precise answers and therefore an optimized user experience.
After entering a user request in the AI Playground, the recognition can be corrected. The feedback is then immediately adopted and influences the recognition of further queries.
In addition to the information on the intent or agent or the result that no assignment was possible, an upward and a downward pointing thumb appear in the top right-hand corner. These symbols can be used to store the query variant as positive or negative feedback.
The AI feedback is entered according to the same principle for agents and editorial intents:
- Confirm correct recognition: If the assignment or recognition is correct, simply click on the upward-facing thumb. This then turns green. The positive feedback can be helpful, as the AI now receives explicit confirmation that the assigned topic is correct.
- If the wrong agent or editorial topic has been recognized, clicking on the downward-facing thumb opens a drop-down menu. The correct topic is selected in the drop-down menu. Just click on the appropriate topic and the feedback is processed when the loading bar has disappeared.
The AI feedback tool is therefore ideal for getting the most out of the AI.
The AI can also be given direct feedback on the recognition of intents and agents in the Insights user requests (hub area statistics, focus reports user requests). This is easily done by clicking on the hat symbol on the right. In this way, queries can be directly fed back to the AI as correct or incorrect during an analysis.