Skip to main content

​
Pattern 1: GraphRAG (Knowledge Graphs + RAG)

Problem: Simple RAG struggles with multi-hop reasoning. Example query: “Which employees who worked on Project X are now working on Project Y?”
  • Requires connecting: Employee → Project X → Employee → Project Y
  • Basic RAG might find docs about each project separately but miss the connection
Solution: Use knowledge graph to find connected entities, then retrieve documents. When to use:
  • Multi-hop reasoning required
  • Entity-centric queries (people, companies, products)
  • Relationship-heavy domains (org charts, supply chains)
Cost: 2-3x more complex than basic RAG Benefit: Often improves multi-hop queries when relationships are explicit in a knowledge graph (magnitude varies)

​
Pattern 2: Iterative RAG (Query Refinement)

Problem: User queries are often vague or poorly formed. Example: User asks “Tell me about the outage”
  • Which outage? (Multiple incidents in database)
  • What aspect? (Cause, impact, resolution, timeline)
Solution: Use LLM to refine query based on initial results. Full runnable example of Iterative RAG When to use:
  • Vague user queries common
  • Large document corpus (many potential matches)
  • Quality > speed (each iteration adds 200-500ms)
Cost: 2-4x basic RAG (multiple retrieval rounds + refinement calls) Benefit: Can improve performance on ambiguous queries by refining intent and terms (magnitude varies)

​
Pattern 3: Agentic RAG (LLM-Driven Retrieval)

Problem: Users don’t know the right keywords or structure. Solution: Let the LLM decide HOW to search based on the question. When to use:
  • Complex, varied query types
  • Large metadata taxonomy (many filter options)
  • Power users who ask sophisticated questions
Cost: 30-50% more than basic RAG (analysis call adds overhead) Benefit: 30-50% improvement on complex queries, better metadata utilization

​
Pattern 4: Hybrid RAG (Structured + Unstructured)

Problem: Some questions need data from both databases AND documents. Example: “What’s our Q4 revenue compared to industry analysis reports?”
  • Revenue: Structured data (database query)
  • Industry analysis: Unstructured data (document retrieval)
When to use:
  • Mixed data sources (databases + documents)
  • Business intelligence + qualitative analysis
  • Compliance (regulations + internal policies)
Cost: Variable (depends on SQL complexity + retrieval) Benefit: Answers questions that pure document RAG can’t handle