Glossary
What is advanced RAG
In short
Advanced RAG is a document search scheme for AI in which the question is refined before the search and the results are filtered and sorted after it.
Search and RAG · advanced RAG
MMOOZZGG.com editorsChecked: 1 min read
For a child
You are looking for a book about space in a library. First you tell the librarian what you want to know: planets or rockets. Then you bring a pile of books, but you do not read them all. You put the extra ones aside and keep two that surely have the answer. Advanced RAG works the same way: it makes the question clear, searches and picks what matters.
In plain words
It is the same “find and answer” scheme, but with two extra steps. Before the search, the system cleans up the question: it rephrases it, adds synonyms and sometimes splits it into several simple ones. After the search, it does not hand the model everything. It sorts the results again and drops repeats. For example, an employee writes “how much leave does a new guy get”. The system rewrites this as “length of annual leave in the first year of work”, finds five pieces in different documents and keeps the two about new employees.
For business
Advanced techniques matter when a prototype already works but makes mistakes on hard questions. It answers from the wrong clause of a contract, mixes up versions of a policy or misses exceptions. Each extra step costs money: query rewriting and reranking add model calls, so the delay and the token bill grow. That is why teams turn techniques on one at a time and check which one really raises the share of correct answers on your own question set. Metadata carries a separate risk: if documents have no dates and version numbers, the system cannot tell a current order from a canceled one. A pilot with a metric agreed in advance shows which techniques pay off for you.
For a specialist
In the classification by Gao and co-authors (2023), advanced RAG keeps the linear pipeline but adds strategies before and after retrieval. Before retrieval (pre-retrieval), teams improve the index: a sliding window when cutting text into chunks, finer segmentation and metadata on each chunk. They also optimize the query: query rewriting, transformation and query expansion. HyDE (Gao, Ma, Lin and Callan, 2022) belongs to this group: the model writes a hypothetical answer document, and the search uses its embedding instead of the embedding of the question. After retrieval (post-retrieval), teams use reranking and context compression. The survey also notes one trick: moving the most relevant passages to the edges of the prompt. Each technique is evaluated on its own: the recall of the retriever before and after, how well the answer is grounded and the added delay.
Example
Lawyers complained that the assistant quoted old versions of contracts. After the team added the version date to the metadata and re-sorted the results, the assistant began to use the current version.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map