How engines work
Retrieval-augmented generation
Also: RAG
Retrieval-augmented generation is when a model searches for relevant documents and writes its answer from what it finds, rather than from training alone. It is why a page published this week can appear in an answer today.
In more detail
RAG is the mechanism behind most citations. The engine retrieves a set of pages, writes an answer grounded in them, and links the ones it used.
It is also the fastest route to influencing an answer: making the right facts retrievable works far quicker than waiting for a future training run.
How Allswap measures this
Guides that cover this
Related terms
Grounding
Grounding is tying a model's answer to specific retrieved sources rather than to its own recollection. A grounded answer can be traced back to documents, which is what makes citations possible and answers more reliable.
Cited source
A cited source is a page or domain that an answer engine links to when answering a question. In most categories a small group of sources accounts for the majority of citations, and that group decides which brands get recommended.
Vector search
Vector search finds documents by comparing embeddings rather than matching keywords, retrieving passages whose meaning is close to the question. It is the retrieval method behind most answer engines.
Large language model
A large language model is a system trained on very large amounts of text to predict and generate language. It is the component that writes an answer engine's response, drawing on what it learned in training plus anything retrieved at the moment of asking.
Ready to Dominate AI Search?
Get started now
Allswap