Courses Blog Research Lab AI Letters The Lab Code Bank Interactive 3DKodr Earnest Jobs

Interactive 3D/Advanced RAG Patterns

HyDE

+12–18% recall

Generate hypothetical doc → embed that → search

Best for: Short or vague queries. The hypothetical document bridges the query–document embedding gap.

User Query

"RLHF training"

→

LLM Hallucinate

Generate fake but plausible answer paragraph

→

Embed Hypothesis

rich hypothetical doc → vector

→

ANN Search

match against real corpus

→

LLM Generate

real context + original query → answer

Why It Works

Dramatically improves recall for short or vague queries. Dense query representation.

Pattern Comparison

Pattern	Extra LLM Calls	Recall Lift	Latency
Naive RAG	0	Baseline	Low
HyDE	1 (gen)	+12–18%	Medium
Parent-Child	0	+9–14%	Low
Step-Back	1 (abstract)	+11–16%	Medium
Multi-Query	1 + N searches	+14–20%	High

Controls

Pattern

Compare

Side-by-side vs Naive RAG

Advanced RAG patterns fix the main failure modes of naive RAG: vague queries (HyDE), missing context (Parent-Child), abstract reasoning (Step-Back), and query ambiguity (Multi-Query). Each adds one LLM call but significantly improves retrieval quality.

Advanced RAG Patterns - Interactive Visualization

Advanced RAG patterns address the failure modes of naive retrieval. HyDE generates a hypothetical answer document to improve query-corpus alignment. Parent-Child retrieval uses small chunks for precision but returns full parent sections for context. Step-Back first retrieves abstract principles before specific facts. Multi-Query expands one query into several variants and unions the results.

HyDE improves recall by 12–18% for short or vague queries by bridging the embedding gap
Parent-Child retrieval combines chunk-level precision with full-section context for the LLM
Step-Back prompting grounds specific answers in broader principles, reducing edge-case hallucination
Multi-Query generation maximizes recall for ambiguous questions with multiple valid interpretations

Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.