ModelRefs / Advanced RAG — AI Glossary
Advanced RAG — AI Glossary
RAG pipelines with pre-retrieval query processing and post-retrieval reranking to improve precision and answer quality. Also called enhanced RAG.
Overview
Advanced RAG adds query rewriting/expansion before retrieval (improves recall), cross-encoder reranking after initial retrieval (improves precision), and context compression before generation (reduces noise). Systems like LlamaIndex and LangChain implement these stages modularly. Addresses the primary failure modes of naive RAG.
Reference details
| Topic | rag |
|---|---|
| Also known as | enhanced RAG |
| Last reviewed | 2026-06-24 |
Related terms
How it works
- Each named stage is an extra model call placed at a specific point in the pipeline, which is why every one of them costs latency before the answer starts.
- Query rewriting runs a model before retrieval to turn a conversational question into one or more retrieval queries, so the search happens on text the corpus would plausibly contain.
- A hypothetical-document variant goes further and generates a fabricated answer, then embeds that rather than the question, on the argument that an answer is closer in vector space to a real answer than a question is.
- Reranking runs after retrieval, scoring each candidate against the query with a model that reads both together, which is why it improves ordering and cannot improve what was never retrieved.
- Compression runs last, either extracting the sentences that bear on the query or summarising the set, trading tokens in the prompt against the risk of discarding the detail the answer needed.
- Each stage acts on a different failure mode and at a different point: recall before retrieval, precision after it, prompt noise before generation — so which stage to add is determined by which of those three measurements is the one that moved.
Example: Fix the stage that is failing
Query rewriting raises recall when users phrase things unlike the documents. Reranking raises precision when the right passage is retrieved but ranked eighth. Compression helps when context is noisy and the model is distracted. Adding all three without measuring means paying three latency costs to fix one problem you have not identified.
Commonly confused with
Advanced RAG is not a product or a fixed architecture — it is a loose label for naive RAG plus optional pre- and post-retrieval stages. Nothing about the term specifies which stages, so it says less about a system than it appears to.
When to use it
Reach for it when:
- Naive RAG is deployed and you have measured where it fails
- You can evaluate each added stage independently
- Latency budget can absorb the extra hops
Reach for something else when:
- Before naive RAG is working and measured — you will not know what helped
- As a default architecture; every stage is latency, cost and another failure mode
- To compensate for bad chunking, which should be fixed at the source
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Advanced RAG — AI Glossary.
Frequently asked questions
What is Advanced RAG?
RAG pipelines with pre-retrieval query processing and post-retrieval reranking to improve precision and answer quality.
Is Advanced RAG the same as enhanced RAG?
Yes — enhanced RAG are common aliases for Advanced RAG.
What concepts are related to Advanced RAG?
Closely related concepts include naive rag, modular rag, query rewriting.