Same philosophy, applied to answers
Why our answers are complete—not confident guesses
The no-guessing rule doesn't stop at automation. When you ask a question about your documents, most tools sample the few passages that look similar to your query and hope the answer is in there. We don't sample.
In simple terms, if information was a huge field of books laying in the dark:
Figurative explanation
Pros
Cons
Plain LLMs would use a single powerful light projector to shine on a large area of the field (called a "context window") and thus illuminate that area (get the data from it).
Able to ingest a lot of data at once. Great for use cases where all the data can be held in the model's context window.
When the model's context window can't hold all the data, it will start "forgetting" things.
The common RAG approach uses powerful lasers to shine on specific areas and thus illuminate them (get the data from them).
Able to reach any point in the data where there is information similar to what the query is talking about.
May miss hidden insights or unknown information if the user isn't sure what to look for exactly.
Our proprietary approach sends numerous small readers, each with a small flashlight, to plow through the entire field bit by bit (not just specific areas). Each reader does one bounded job and reports back—no reader decides anything on its own.
Able to find any and all hidden insights in the information, relevant to the query, either explicitly or implicitly.
Slower. It may take a few minutes to get the results, depending on the size of the data. However, there are ways to speed it up (see smart snapshots).