Skip to content

EdgeQuake vs Traditional RAG

Why Knowledge Graphs Transform Retrieval Quality

Traditional RAG (Retrieval-Augmented Generation) uses vector similarity search to find relevant document chunks. EdgeQuake adds knowledge graph construction, enabling semantic understanding of entity relationships that pure vector search misses.


AspectTraditional RAGGraph-Enhanced RAG (EdgeQuake)
RetrievalVector similarity onlyVector + Graph traversal
UnderstandingSemantic similarityEntity relationships
Multi-hop❌ Single-hop✅ Multi-hop reasoning
Themes❌ Local only✅ Global themes
IndexingFast (~1s/doc)Slower (~5-30s/doc)
Query Latency~100-300ms~200-500ms

Traditional RAG has fundamental limitations:

Consider this document:

“Sarah Chen works at MIT. She authored the climate paper with Dr. James Wilson.”

Question: “What is the connection between Sarah Chen and James Wilson?“

┌─────────────────────────────────────────────────────────────────┐
│ TRADITIONAL RAG PROBLEM │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Document chunks: │
│ ┌────────────────────────────────────────┐ │
│ │ Chunk 1: "Sarah Chen works at MIT..." │ → embedding_1 │
│ └────────────────────────────────────────┘ │
│ ┌────────────────────────────────────────┐ │
│ │ Chunk 2: "She authored the climate..." │ → embedding_2 │
│ └────────────────────────────────────────┘ │
│ │
│ Query: "connection between Sarah and James" │
│ │
│ Vector search: May find Chunk 1 (Sarah mentioned) │
│ May miss Chunk 2 (if "connection" not similar) │
│ │
│ PROBLEM: No explicit link between Sarah and James! │
│ │
└─────────────────────────────────────────────────────────────────┘

Question: “What are the main themes in this 50-page document?”

Traditional RAG retrieves the most semantically similar chunks to “themes”, but this misses the document’s structure and organization.

Question: “Who are Sarah Chen’s collaborators’ organizations?”

This requires:

  1. Find Sarah Chen
  2. Find her collaborators
  3. Find their organizations

Vector search cannot chain these lookups.


EdgeQuake constructs a knowledge graph during indexing:

┌─────────────────────────────────────────────────────────────────┐
│ GRAPH-ENHANCED RAG │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Document → LLM Extraction → Knowledge Graph │
│ │
│ ┌───────────────┐ │
│ │ SARAH_CHEN │ │
│ │ (PERSON) │ │
│ └───────┬───────┘ │
│ │ │
│ ┌───────────┼───────────┐ │
│ │ WORKS_AT │ CO_AUTHORED │
│ ▼ ▼ │
│ ┌─────┐ ┌──────────────┐ │
│ │ MIT │ │ CLIMATE_PAPER│ │
│ └─────┘ └──────┬───────┘ │
│ │ AUTHORED_BY │
│ ▼ │
│ ┌──────────────┐ │
│ │ JAMES_WILSON │ │
│ │ (PERSON) │ │
│ └──────────────┘ │
│ │
│ Query: "connection between Sarah and James" │
│ │
│ Graph traversal: SARAH_CHEN → CLIMATE_PAPER → JAMES_WILSON │
│ Relationship: CO_AUTHORED │
│ │
│ ANSWER: "Sarah and James co-authored the climate paper" │
│ │
└─────────────────────────────────────────────────────────────────┘

FeatureTraditional RAGEdgeQuake
Chunk embedding
Entity extraction
Relationship extraction
Knowledge graph
Multi-hop queries
Theme detection
Entity-centric search
Global context
Source lineage⚠️ Basic✅ Full

Research from the LightRAG paper (arxiv:2410.05779) shows significant improvements:

DatasetTraditional RAGGraph-RAGImprovement
Agriculture32.4%67.6%+35%
CS38.4%61.6%+23%
Legal16.4%83.6%+67%
Mix38.8%61.2%+22%

Metrics: Comprehensiveness, measured by LLM-as-judge evaluation


Graph-enhanced RAG requires more processing at index time:

┌─────────────────────────────────────────────────────────────────┐
│ INDEXING COMPARISON │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Traditional RAG: │
│ ┌────────┐ ┌───────────┐ ┌─────────────┐ │
│ │ Doc │ ─▶ │ Chunk │ ─▶ │ Embed │ ─▶ Done │
│ │ │ │ (~10ms) │ │ (~100ms) │ │
│ └────────┘ └───────────┘ └─────────────┘ │
│ │
│ Total: ~200ms per document │
│ │
│ ────────────────────────────────────────────────────────────── │
│ │
│ EdgeQuake: │
│ ┌────────┐ ┌───────────┐ ┌─────────────┐ │
│ │ Doc │ ─▶ │ Chunk │ ─▶ │ LLM Extract │ ─▶ ─┐ │
│ │ │ │ (~10ms) │ │ (~2-10s) │ │ │
│ └────────┘ └───────────┘ └─────────────┘ │ │
│ ▼ │
│ ┌─────────────┐ │
│ │ Graph Merge │ │
│ │ (~100ms) │ │
│ └──────┬──────┘ │
│ ▼ │
│ ┌─────────────┐ │
│ │ Embed │ │
│ │ (~200ms) │ │
│ └─────────────┘ │
│ │
│ Total: ~5-30s per document │
│ │
└─────────────────────────────────────────────────────────────────┘
ScenarioTraditional RAGEdgeQuakeRecommendation
100 docs, simple queries✅ Fast, cheapOverkillTraditional
100 docs, relationship queries❌ Poor quality✅ GoodEdgeQuake
10K docs, mixed queriesFast, moderate qualitySlower index, better qualityEdgeQuake
Real-time indexing needed✅ Works⚠️ LatencyTraditional

  • ✅ Documents have simple, factual content
  • ✅ Queries are keyword-based lookups
  • ✅ Real-time indexing is required
  • ✅ LLM costs are a primary concern
  • ✅ You need minimal infrastructure
  • ✅ Documents describe entities and relationships
  • ✅ Users ask about connections and themes
  • ✅ Multi-hop reasoning is needed
  • ✅ Answer quality is more important than indexing speed
  • ✅ Global document understanding is required

EdgeQuake’s query modes let you blend both approaches:

ModeStrategyUse Case
naiveVector onlySimple factual queries
localVector + Entity graphEntity-specific questions
globalGraph communitiesTheme/overview questions
hybridAll approachesComplex queries (default)

This means you get the best of both worlds:

  • Fast vector search for simple queries
  • Graph traversal for complex reasoning
  • Combined context for comprehensive answers

AspectTraditional RAGEdgeQuake
Setup complexityLowMedium
LLM calls per doc1 (embedding)3-10 (extraction + embedding)
InfrastructureVector DB onlyVector + Graph DB
MaintenanceSimpleModerate
Query tuningLimited6 modes to optimize

┌─────────────────────────────────────────────────────────────────┐
│ DECISION MATRIX │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Question Type │ Traditional │ EdgeQuake │
│ ───────────────────────────────────────────────────────────── │
│ "What is X?" │ ⭐⭐⭐ ⭐⭐⭐
│ "How does X work?" │ ⭐⭐ ⭐⭐⭐
│ "What connects X and Y?" │ ⭐ ⭐⭐⭐⭐
│ "Main themes in doc?" │ ⭐ ⭐⭐⭐⭐
│ "X's collaborators' orgs?" │ ❌ ⭐⭐⭐⭐
│ If most queries are multi-hop or relationship-based, │
│ EdgeQuake provides significantly better results. │
│ │
└─────────────────────────────────────────────────────────────────┘