Maven quality
How to think about Maven quality
Maven is designed to answer questions with expert human quality. A high quality answer should be:
- Factually correct, complete, and helpful
- Empathetic, personalized, and with a human touch
- Timely
Understanding the quality calculation
To compute the quality of a response, Maven uses a combination of signals including
- Automatic self grading: Maven has a self-grading mechanism that displays how confident it is based on how much relevant information it has in knowledge.
- Automatic grading against human responses: Maven automatically grades historic tickets against human responses where human agent responses are available
- Human feedback from Copilot and end-user thumbs up and down
Improving quality
Answers from Maven are only as good as the knowledge, user data, and personalization that has been provided.
When Maven gives a low-quality answer, there are several common reasons why:
Missing knowledge
It’s best to think of Maven as a very junior support agent. Eager to help, but doesn’t know much about your company or product. So sometimes, Maven is unable to answer a question because it simply does not have the information it needs to do so.
To see if this is the case, open up the Conversation details page on a low performing question and click on the “View Details” button underneath the Maven response. This section shows the knowledge documents and actions Maven used in crafting its response.
If the knowledge listed does not include the answer within it, then Maven won’t be able to do a good job. Either:
- the knowledge needed is not in your knowledge base
- or the knowledge is available, but it wasn’t a good match for the user question
For the first, simple add the missing knowledge to your knowledge base. Create a “Manual” knowledge base within Maven if you don’t already have one. Then we recommend adding a document with the title set to the user’s question and the content set your preferred answer.
For the second, it’s helpful to know that Maven does a search of the knowledge base for each customer question it receives. If the knowledge document with the answer didn’t show up as a search result, then either it’s not specific enough for the user question, or if it’s from a URL knowledge base, it may be low quality.
As a quick fix, you can add duplicate knowledge using the above solution for the first issue. It may also be helpful though to double check all of your URL knowledge to make sure that the resulting content looks reasonable. The more readable it is to a human, the better it will be for search results and the LLM.
Conflicting information
As described above, your junior support agent is always eager, but if you give them two conflicting pieces of information, they won’t really know what to do! They will probably get their wires crossed and give all sorts of incorrect answers.
Also as above, looking at the knowledge used for a low performing question can help you identify the issue. To see if this is the case, open up the Conversation details page and click on the “View Details” button underneath the Maven response.
The exact paragraphs sent to the LLM will be shown in this section. If these paragraphs are conflicting or even just a bit ambiguous, then it’s worth fixing the knowledge documents used to answer the question.
If it’s not possible to fix the source documents on your help center or website, then it’s always possible to deactivate single knowledge documents within Maven. Then, once the bad information is removed, knowledge can be manually added within the dashboard and used until the source documents are fixed.
Off-topic answers
Remember that Maven doesn’t know anything about your company that it has not been told.
It’s highly recommended that a basic overview of your company or product be added within the persona instructions. What is your company name? What do you offer or sell? And what similar services do you not provide?
During your onboarding process, Maven will help you specify these key details.