SmartSearch is built using a technique called Retrieval-Augmented Generation (RAG). This means all uploaded documents are read and split into small sections of text. Each section is then interpreted by a large language model and stored in something called a vector database. This is a database that is built specifically to map the meaning of text to coordinates in a multi-dimensional space. Sounds complicated right?

Think of online shops instead of files and how they suggest similar products to you when you shop. They do the same thing: they take the product you are looking at, search in a vector database and find similar products to the one you are looking at. SmartSearch does the same with text — your question is interpreted and matched against the parts in your files that are most similar. Those parts are handed to the AI as context, along with a record of exactly which document and paragraph they came from. That record allows us to show citations from your files, so you know exactly where the information came from and can verify with one click.

This approach also makes it really easy to find specific information across hundreds of files, even if you don't know the exact terminology. Since we are using an LLM to interpret your question, you can just describe what you are looking for and SmartSearch will find it. Even if it is hidden in hundreds of large and complicated files.

A big advantage to this is also almost no hallucinations. Since every factual statement must be backed by a citation, the AI will tell you when the information simply isn't in your documents. Instead of making it up Smartsearch informs you that the context is missing and more information is needed.

So in short the process has four critical steps to it: