v2.4Sign In
RAG Types
Graph RAG — Retrieval Pipeline
How queries are processed with vector + graph traversal.
Overview
Graph RAG retrieval combines vector search with knowledge graph traversal, enabling multi-hop reasoning and relationship-based queries that pure vector search cannot handle.
Pipeline Steps
1. Query Embedding
The query is embedded for vector search using the same model as indexing.
2. Graph Query Planning
Analyzes the query to identify entities and relationships for graph traversal.
3. Dual Search
- Vector Search: Finds semantically similar chunks
- Graph Traversal: Navigates the knowledge graph from query entities
4. Graph Traversal
| Step | Description |
|---|---|
| Entity Lookup | Identifies entities mentioned in the query |
| Relationship Path Finding | Finds paths between entities |
| Multi-Hop Navigation | Traverses multiple relationship hops |
5. Result Fusion
Combines vector scores with graph path relevance for unified ranking.
6. Result Ranking
Final ranking based on fused scores.
7. Context Assembly
Combines top chunks into context, including graph-derived insights.
8. Response Generation (Optional)
- Activated when:
llmEnabled = true - What it does: Passes assembled context + query to LLM
Key Differences from Hybrid RAG
- Graph Traversal replaces BM25 as second retrieval path
- Graph Query Planning identifies entities in query for traversal
- Result Fusion combines vector scores + graph path relevance
- Supports multi-hop queries ("Who manages the team that built X?")