SmartFAQs.ai
Back to Learn
Deep Dive

Reflective RAG

Reflective RAG is an architectural pattern that introduces self-correction loops into the retrieval-augmented generation process, where an LLM critiques its own retrieved context for relevance and its generated response for factual consistency. This iterative approach significantly reduces hallucinations by discarding 'noisy' documents and re-triggering retrieval or generation if quality thresholds are not met, trading increased latency and token cost for high-precision output.

Definition

Reflective RAG is an architectural pattern that introduces self-correction loops into the retrieval-augmented generation process, where an LLM critiques its own retrieved context for relevance and its generated response for factual consistency. This iterative approach significantly reduces hallucinations by discarding 'noisy' documents and re-triggering retrieval or generation if quality thresholds are not met, trading increased latency and token cost for high-precision output.

Disambiguation

Distinguished from 'Naive RAG' by its ability to recursively evaluate and fix its own errors before presenting a final answer.

Visual Metaphor

"A research student who writes a draft, compares it against their source notes to find contradictions, and then re-writes the draft or visits the library again if information is missing."

Key Tools
LangGraphLlamaIndex WorkflowsDSPyLangChain Expression Language (LCEL)
Related Connections

Conceptual Overview

Reflective RAG is an architectural pattern that introduces self-correction loops into the retrieval-augmented generation process, where an LLM critiques its own retrieved context for relevance and its generated response for factual consistency. This iterative approach significantly reduces hallucinations by discarding 'noisy' documents and re-triggering retrieval or generation if quality thresholds are not met, trading increased latency and token cost for high-precision output.

Disambiguation

Distinguished from 'Naive RAG' by its ability to recursively evaluate and fix its own errors before presenting a final answer.

Visual Analog

A research student who writes a draft, compares it against their source notes to find contradictions, and then re-writes the draft or visits the library again if information is missing.

Related Articles