To reduce the amount of questions that you actually have to answer manually we've given our Slack bot the superpowers (and authority) to answer questions for you.
But how does that really work? In this article we're presenting both the short and the long answer to this question, as well as tips and tricks on how to help the algorithm make better suggestions.
The short answer
Every answer in the Knowledge is analyzed by our algorithm and included in the model.
Incoming questions on Slack, MS Teams, or Google Chat are compared to the model.
If there is a good enough match, fitting answers are suggested to the employee.
Over time the model learns from employee feedback to provide better and better suggestions.
The longer answer
Everything starts with the answers you add in the Knowledge section. The suggestion algorithm works by using Natural Language Processing to ingest every answer in your Knowledge section and to extract the most important topics.
Whenever you add a new answer, Back immediately breaks down the text (title and contents) into phrases and transforms it into a numeric vector. In other words, towards our bot the answer is represented not as text, but rather as a list of numbers, which can be interpreted as the answer's "score" across different topics. For instance, you can think of the examples:
What is the office-ness score of this answer on a scale from -1 to 1?
What is the invoice-ness score of this answer on a scale from -1 to 1?
When an employee asks the Back bot on Slack, MS Teams or Google Chat a question, it is also automatically broken down into a numeric vector. All the answers in your organization are then scored based on how closely their vectors match the vector of the question.
If some of the answers are similar enough to the question, we show them as suggestions to the employee (max. 3 answers are suggested).
Such example is visualized below - in reality, each answer is evaluated and compared on as many dimensions as there are in the incoming question.
Learning from feedback
Over time, our algorithm learns from user feedback and suggests better answers.
Whenever an employee receives a suggestion, they can provide feedback whether the suggestion was relevant or not using the two buttons ✅ This solves it or I need help from an expert.
When another employee asks a similar question, the algorithm will retrieve all the answers that were suggested to all the similar questions and take their feedback into account.
Answers, that were marked as helpful will then receive a small boost in the ranking. Those, that were dismissed as irrelevant, will receive a small penalty.
In a nutshell, the more your employees interact with the Back bot and the more feedback they give it, the better it will learn to surface the correct answers from your Knowledge Base.
How to help the algorithm make great suggestions
💁🏻♀️ Write as your employees do.
Short, concise answers with a crisp, descriptive title make for the best suggestions. If you want your answer to be easily found, make sure to write them in the same kind of language that your employees would naturally use to ask a question on the topic. If your employees tend to use various words to describe the same topic (e.g. "vacation" or "holidays"), you can make them easier to find by including both possible phrasings in your answers.
🤖 Only add answers you're happy to suggest automatically.
If you don't want an answer to be suggested to users (for example because it includes sensitive information), consider adding it as a Saved reply instead. Those can only be used by you directly when answering the employee through a private conversation.
🤹♂️ Try it out yourself!
Once you add a new answer to the Knowledge Base, it is ready to be suggested within 1 minute. Try to ask the Back bot a question that you know should result in this article being suggested and give the bot some feedback to steer it in the right direction.
Some things to keep in mind:
Currently we only index answers and display suggestions written in English. Answers written in other languages will not be automatically suggested (although you can still attach them to replies manually).
We don't show suggestions for very short queries (1-3 words), as they usually don't contain enough information to understand the topic.
User feedback is updated every minute, so you might not see the answers change immediately after giving feedback. Additionally, we find that positive feedback towards a suggestion is a much stronger signal than a negative one, so we value positive feedback more than negative feedback. Negative feedback will not make your documents disappear and never be suggested again. It just makes the document slightly less likely to resurface for a similar question in the future.
Both answers and employee questions are anonymized before making any suggestions, to ensure no sensitive data is "learned" by the algorithm.