Decision Guide

Use this guide to identify which patterns are most relevant to your current situation. Start with the question that best describes your primary concern.

I need to reduce costs

How are you currently calling LLMs?

• Direct API calls with no intermediary → Start with LLM Gateway Pattern
• Single provider, single model for everything → Start with Model Router Pattern
• Seeing many repeated or similar queries → Start with Semantic Caching
• Long context windows eating your budget → Start with Token Budget Pattern
• Most queries are simple but you pay for maximum context on every request → Start with Cascading Context Assembly

I need to improve reliability

What kind of failures are you seeing?

• Provider outages causing downtime → Start with Circuit Breaker for LLMs
• Quality degradation after prompt changes → Start with Prompt Canary Deployment and LLM-as-Judge
• Silent quality drops nobody notices → Start with Embedding Drift Detector and Span-Level Tracing

I need better retrieval / RAG

What does your knowledge base look like?

• Complex relational data, multi-hop questions → GraphRAG Pattern
• Need both keyword and semantic matching → Hybrid Search Pattern
• Retrieval returns stale or outdated content → Retrieval Freshness Watermark
• Retrieval returns redundant/duplicate chunks → Semantic Deduplication
• Upstream data keeps breaking your pipeline → Data Contract Pattern

I need to handle security and compliance

What is your primary concern?

• Prompt injection and jailbreak attempts → Input Sanitization Pattern
• Agents calling external tools that return untrusted data → Tool Output Firewall
• Need standardized model documentation for compliance → Model Card Pattern

I need observability

What can you not see today?

• What is happening inside multi-step chains → Span-Level Tracing Pattern
• Whether retrieval quality is silently degrading → Embedding Drift Detector

I am building agent systems

Agents combine multiple patterns. A typical production agent stack includes:

1. LLM Gateway Pattern for routing and observability
2. Input Sanitization for the front door
3. Tool Output Firewall for the side door (tool results re-entering context)
4. Circuit Breaker for LLMs for provider resilience
5. Token Budget Pattern for runaway prevention
6. Span-Level Tracing for debugging multi-step flows

I need graph-based intelligence

What are you trying to achieve with graphs?

• Multi-hop questions that connect information across documents → Start with GraphRAG
• Non-linear reasoning with branching and merging approaches → Start with Graph of Thoughts
• Deduplicating entities across multiple data sources → Start with Entity Resolution Graph
• Building a knowledge graph for your domain → Start with GraphRAG for the retrieval layer and Entity Resolution Graph for clean entity data

I need to evaluate and test AI quality

What is your primary evaluation challenge?

• Need automated quality scoring at scale → Start with LLM-as-Judge
• Quality is degrading after prompt or model changes → Combine LLM-as-Judge with Prompt Canary Deployment
• Need a quality gate before deploying prompt changes → Use Prompt Canary Deployment with LLM-as-Judge as the scoring mechanism