How does Copilot come up with answers?

Copilot is an Artificial Intelligence (AI) tool, and its foundation is a Large Language Model (LLM). It is capable of reading a question asked in natural language, like the kind of question you might ask a coworker. It references certain information to respond using natural language, but a lot happens in between asking your question and receiving Copilot's response. Here is how Copilot works behind the scenes:

1. First, you ask a question in natural language. As opposed to the search engine technology we're used to, Copilot responds much better to questions asked using full sentences, instead of just a few keywords. It needs your full thought written out to interpret the intent of what you're asking, just like a person might. It then takes your natural language question, and converts it into a sequence of numerical values on a kind of map, called vectors.
2. Next, Copilot determines the intent of your question. When we talk about interpreting intent, we mean that Copilot is deciding which "skill", or database, to access to find the right information to answer your question. If your intent is to find out how to do something in Procore, Copilot would send your query in mathematical form over to the skill that addresses all questions about how to use Procore. The skill looks at the numerical values from your question, and finds the closest values present in the database to determine which information should be used to answer your question.
3. Then, information is sent from the service back to Copilot.  When Copilot receives the information from the relevant skill, still in numerical format, it collects it and translates it back to natural language. Finally, Copilot presents the answer to you along with links to the files or items in Procore where it found the answer to your question.